sentences
sequence | new_gt
sequence |
---|---|
[
"1",
")",
"We",
"present",
"CLPM",
",",
"a",
"dynamic",
"and",
"token",
"-",
"wise",
"masking",
"scheme",
"using",
"special",
"tokens",
"[",
"C",
"]",
"x",
",",
"to",
"form",
"cross-",
"lingual",
"forward",
"passes",
"in",
"multilingual",
"pre",
"-",
"training"
] | [
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName"
] |
[
"[",
"C",
"]",
"xis",
"a",
"generalized",
"representation",
"from",
"multiple",
"cross",
"-",
"lingual",
"candidates",
"2",
")",
"CLPM",
"substantially",
"improves",
"the",
"performance",
"of",
"X↔Enbaseline",
"UNMT",
"models",
"by",
"3",
"%",
"∼8",
"%",
"on",
"rich",
"-",
"resource",
"and",
"low",
"-",
"resource",
"languages",
"and",
"can",
"facilitate",
"training",
"on",
"dissimilar",
"languages"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"3",
")",
"Beyond",
"UNMT",
"tasks",
"or",
"bilingual",
"tasks",
",",
"CLPM",
"can",
"be",
"used",
"for",
"cross-",
"lingual",
"classification",
"tasks"
] | [
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"B-TaskName",
"I-TaskName",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName"
] |
[
"2",
"Cross",
"-",
"lingual",
"Prototype",
"Masking",
"Notation",
"Lxis",
"the",
"language",
"ID",
"of",
"language",
"Lang",
"x.",
"Pnstands",
"for",
"positions"
] | [
"O",
"B-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"ERis",
"the",
"embed-",
"ding",
"for",
"R.dis",
"the",
"model",
"/",
"embedding",
"dimension.2.1",
"Forward",
"Pass",
"in",
"Attention",
"Given",
"an",
"input",
"sentence",
"X=",
"{",
"x0",
",",
"x1",
",",
"...",
",",
"x",
"n",
"}",
"in",
"the",
"language",
"Lang",
"x",
",",
"the",
"self",
"-",
"attention",
"layer",
"(",
"Vaswani",
"et",
"al",
",",
"2017",
")",
"performs",
"on",
"the",
"sum",
"of",
"Xinput",
"=",
"{",
"Ex0+ELx+EP0",
",",
"...",
",",
"E",
"xn+ELx+",
"EPn",
"}",
",",
"which",
"is",
"considered",
"in",
"previous",
"works",
"of",
"multilingual",
"pre",
"-",
"training",
"(",
"Liu",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"O",
"O"
] |
[
",",
"2020",
";",
"Song",
"et",
"al",
",",
"2019",
";",
"Lample",
"and",
"Conneau",
",",
"2019",
")",
"For",
"pre-",
"dicting",
"xi",
",",
"the",
"attention",
"score",
"(",
"Bahdanau",
"et",
"al",
",",
"2015",
";",
"Luong",
"et",
"al",
",",
"2015",
")",
"ei",
",",
"j=",
"(",
"Exi+ELx+",
"EPi",
")",
"TWT",
"qWk",
"(",
"Exj+ELx+EPj",
")",
"between",
"query",
"vector",
"qiand",
"key",
"vector",
"kjwithin",
"the",
"same",
"sen-",
"tence",
"can",
"be",
"decomposed",
":",
"ei",
",",
"j",
"=",
"ET",
"xiWT",
"qWkExj",
"/",
"bracehtipupleft",
"/",
"bracehtipdownright",
"/",
"bracehtipdownleft",
"/",
"bracehtipupright",
"a+",
"ELx",
"(",
"·",
")",
"/",
"bracehtipupleft",
"/",
"bracehtipdownright",
"/",
"bracehtipdownleft",
"/",
"bracehtipupright",
"b+EPi",
"(",
"·",
")",
"/",
"bracehtipupleft",
"/",
"bracehtipdownright",
"/",
"bracehtipdownleft",
"/",
"bracehtipupright",
"c+EPj",
"(",
"·",
")",
"/",
"bracehtipupleft",
"/",
"bracehtipdownright",
"/",
"bracehtipdownleft",
"/",
"bracehtipupright",
"d",
"(",
"1",
")",
"where",
"WqandWkare",
"linear",
"transformation",
"for",
"the",
"query",
"vector",
"qiand",
"key",
"vector",
"kjrespectively",
",",
"andiandjstands",
"for",
"position",
"indexes",
"Terms",
"(",
"b",
")",
",",
"(",
"c",
")",
",",
"and",
"(",
"d",
")",
"introduce",
"the",
"inductive",
"bias",
"towards",
"language",
"Lang",
"x",
",",
"position",
"Pi",
",",
"and",
"position",
"Pjre-",
"spectively",
"When",
"predicting",
"xi",
",",
"we",
"have",
"the",
"for-",
"ward",
"pass",
":",
"{",
"xi",
",",
"xj\\i",
"}",
"→xi",
",",
"where",
"xj\\idenotes",
"all",
"the",
"tokens",
"around",
"position",
"i",
",",
"and",
"the",
"prediction",
"of",
"xiis",
"conditioned",
"by",
"{",
"xi",
",",
"xj\\i",
"}",
"The",
"forward",
"pass",
"ismonolingual",
"because",
"both",
"two",
"sides",
"are",
"in",
"the",
"same",
"language",
"In",
"optimization",
",",
"we",
"can",
"compute",
"gradients",
"from",
"the",
"backward",
"pass",
":",
"∂εxi",
"∂Exiand∂εxi",
"∂Exj",
",",
"where",
"εxiis",
"the",
"predicting",
"error",
"2.2",
"MLM",
"with",
"[",
"M",
"]",
"and",
"CBOW",
"Suppose",
"xiis",
"randomly",
"selected",
"to",
"be",
"replaced",
"by",
"[",
"M",
"]",
"Term",
"(",
"a",
")",
"is",
"changed",
"to",
"ET",
"[",
"M",
"]",
"WT",
"qWkExj",
"Since",
"[",
"M",
"]",
"does",
"not",
"provide",
"prior",
"information",
"of",
"xi",
",",
"Term",
"(",
"a",
")",
"forms",
"a",
"built",
"-",
"in",
"CBOW1model",
"(",
"Continu-",
"ous",
"Bag",
"-",
"of",
"-",
"Words",
"(",
"Mikolov",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"O",
"O",
"O",
"O"
] |
[
",",
"2013",
")",
")",
"learning",
"CBOW",
"or",
"bidirectional",
"information"
] | [
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O"
] |
[
"The",
"forward",
"pass",
"{",
"[",
"M",
"]",
",",
"xj\\i",
"}",
"→xiis",
"still",
"monolingual",
"in",
"mul-",
"tilingual",
"pre",
"-",
"training",
"because",
"[",
"M",
"]",
"is",
"shared",
"and",
"agnostic",
"for",
"all",
"the",
"languages"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"However",
",",
"the",
"model",
"is",
"significantly",
"encouraged",
"to",
"predict",
"xiby",
"under-",
"standing",
"neighboring",
"tokens",
"xj\\iin",
"the",
"sentence",
",",
"i.e.",
",",
"the",
"surrounding",
"context",
"or",
"bidirectional",
"infor-",
"mation",
"Moreover",
",",
"since",
"[",
"M",
"]",
"is",
"overlapping",
"and",
"1For",
"instance",
",",
"given",
"X=",
"{",
"x0",
",",
"[",
"M",
"]",
",",
"x2",
",",
"x3",
"}",
",",
"we",
"have",
"the",
"forward",
"pass",
":",
"{",
"xi=",
"[",
"M",
"]",
",",
"xj\\i=",
"(",
"x0",
",",
"x1",
",",
"x3",
")",
"}",
"→x2",
"if",
"predicting",
"x2",
",",
"where",
"{",
"xi=",
"[",
"M",
"]",
",",
"xj\\i=",
"(",
"x0",
",",
"x1",
",",
"x3",
")",
"}",
"models",
"(",
"non",
"-",
"standard",
")",
"CBOW",
"(",
"4",
"-",
"gram",
")",
".856",
"shared",
",",
"and",
"xj\\iare",
"potentially",
"overlapping",
"tokens",
"in",
"different",
"languages",
",",
"it",
"refines",
"the",
"morphology",
"of",
"different",
"languages",
"to",
"overlap",
"each",
"other",
"for",
"form-",
"ing",
"the",
"isomorphic",
"spaces",
"(",
"Karthikeyan",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
",",
"2020",
";",
"Wu",
"and",
"Dredze",
",",
"2019",
";",
"Pires",
"et",
"al",
",",
"2019",
";",
"Dufter",
"and",
"Schütze",
",",
"2020",
")",
"and",
"leverages",
"domain",
"adaptation",
"(",
"Ganin",
"et",
"al",
",",
"2016",
")",
"or",
"language",
"adaptation",
"(",
"Ai",
"and",
"Fang",
",",
"2022b",
")",
"2.3",
"MLM",
"with",
"[",
"C",
"]",
"x",
"Although",
"the",
"forward",
"pass",
"{",
"[",
"M",
"]",
",",
"xj\\i",
"}",
"→xisig-",
"nificantly",
"enables",
"the",
"model",
"to",
"learn",
"both",
"cross-",
"lingual",
"and",
"monolingual",
"knowledge",
"from",
"the",
"shared",
"token",
"[",
"M",
"]",
"(",
"Dufter",
"and",
"Schütze",
",",
"2020",
")",
"and",
"structural",
"information",
"of",
"the",
"neighboring",
"tokens",
"xj\\i",
"(",
"Karthikeyan",
"et",
"al",
",",
"2020",
";",
"Pires",
"et",
"al",
",",
"2019",
")",
"in",
"multilingual",
"MLM",
"pre",
"-",
"training",
",",
"learning",
"cross-",
"linguality",
"is",
"implicit",
"and",
"limited"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Our",
"idea",
"is",
",",
"we",
"can",
"replace",
"[",
"M",
"]",
"withxi",
"’s",
"cross",
"-",
"lingual",
"prototype",
"[",
"C",
"]",
"xithat",
"we",
"explicitly",
"have",
"a",
"principled",
"cross-",
"lingual",
"forward",
"pass",
":",
"{",
"[",
"C",
"]",
"xi",
",",
"xj\\i",
"}",
"→",
"xi",
"In",
"this",
"way",
",",
"we",
"inject",
"weak",
"but",
"explicit",
"cross",
"-",
"lingual",
"super-",
"vision",
"into",
"the",
"model",
"in",
"multilingual",
"pre",
"-",
"training"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName"
] |
[
"Therefore",
",",
"we",
"replace",
"the",
"selected",
"xiwith",
"its",
"[",
"C",
"]",
"xi",
"instead",
"of",
"[",
"M",
"]",
"as",
"presented",
"in",
"the",
"example",
"(",
"Table",
"1",
")",
",",
"and",
"Term",
"(",
"a",
")",
"is",
"modified",
"to",
"ET",
"[",
"C",
"]",
"xiWT",
"qWkExj",
"accordingly",
"2.4",
"On",
"-",
"the",
"-",
"fly",
"[",
"C",
"]",
"x",
"To",
"obtain",
"[",
"C",
"]",
"xiwithout",
"any",
"cross",
"-",
"lingual",
"supervi-",
"sion",
"in",
"multilingual",
"pre",
"-",
"training",
",",
"the",
"starting",
"point",
"is",
"the",
"output",
"distribution",
"over",
"the",
"vocabulary",
"V",
"shared",
"by",
"all",
"the",
"languages",
"Given",
"the",
"multilin-",
"gual",
"model",
"Net",
",",
"we",
"set",
"Net",
"tothe",
"inference",
"mode",
",",
"not",
"the",
"MLM",
"pre",
"-",
"training",
"mode",
",",
"and",
"the",
"proba-",
"bility",
"of",
"xiis",
"obtained",
"from",
"the",
"softmax",
"layer",
"Qxi",
"=",
"exp",
"(",
"hT",
"xi",
"&",
"LxOxi",
")",
"/",
"summationtextV",
"k=1exp",
"(",
"hT",
"xi",
"&",
"LxOxk",
")",
",",
"where",
"hxi",
"&",
"Lx∈",
"Net",
"(",
"Ex+ELx",
")",
"is",
"the",
"contextualized",
"representa-",
"tion",
"of",
"xi",
",",
"Ex=",
"{",
"Ex0",
",",
"Ex1",
",",
",",
"E",
"xn",
"}",
"is",
"the",
"em-",
"bedding",
"of",
"the",
"input",
"sentence",
",",
"and",
"Oxis",
"factorized",
"from",
"the",
"output",
"matrix2O",
"Recall",
"that",
",",
"in",
"Eq",
"1",
",",
"the",
"language",
"embedding",
"ELxof",
"the",
"language",
"Lang",
"x",
"associated",
"with",
"the",
"token",
"xintroduces",
"inductive",
"bias",
"towards",
"Lang",
"x",
",",
"so",
"that",
"hxi",
"&",
"Lxis",
"biased",
"by",
"ELxtowards",
"Lang",
"xand",
"generalized",
"from",
"Exi",
"In",
"this",
"way",
",",
"the",
"output",
"distribution",
"over",
"the",
"vocab-",
"ulary",
"is",
"biased",
"by",
"ELxtowards",
"Lang",
"x",
",",
"and",
"the",
"dot",
"-",
"products",
"distinguish",
"relevant",
"tokens",
"from",
"irrel-",
"evant",
"tokens",
"for",
"xi",
"Intuitively",
",",
"we",
"can",
"fool",
"the",
"2Note",
"that",
",",
"in",
"most",
"of",
"the",
"cases",
",",
"the",
"output",
"matrix",
"shares",
"all",
"the",
"parameters",
"with",
"the",
"embedding",
"matrix.model",
"by",
"inputting",
"Ex+ELy3",
"The",
"result",
"is",
"that",
"hxi",
"&",
"Ly∈Net",
"(",
"Ex+ELy",
")",
"is",
"biased",
"by",
"ELyto-",
"wards",
"Lang",
"ybut",
"still",
"generalized",
"from",
"Exi",
"We",
"expect",
"hxi",
"&",
"Lyto",
"be",
"an",
"agnostic",
"representation",
"that",
"is",
"relevant",
"to",
"xiandLang",
"y",
"Then",
",",
"we",
"can",
"factorize",
"Oyfrom",
"the",
"output",
"matrix",
"and",
"rank",
"the",
"dot",
"prod-",
"ucthT",
"xi",
"&",
"LyOyto",
"search",
"relevant",
"tokens",
"for",
"xiin",
"Lang",
"yfrom",
"the",
"output",
"space",
"We",
"will",
"discuss",
"the",
"inspiration",
"later",
",",
"and",
"in",
"our",
"experiment",
",",
"we",
"show",
"a",
"case",
"study",
"that",
"some",
"useful",
"candidates",
"in",
"the",
"other",
"language",
"are",
"obtained",
"We",
"approximate",
"a",
"relevant",
"candidate",
"set",
"PY",
"xiin",
"the",
"other",
"language",
"Lang",
"yand",
"compute",
"a",
"weighted",
"average",
"of",
"candidates",
"’",
"embeddings",
",",
"where",
"PY",
"xicon-",
"tributes",
"to",
"low",
"variance",
"and",
"rich",
"information",
"For-",
"mally",
",",
"we",
"define",
"E",
"[",
"C",
"]",
"xi=",
"/",
"summationtext",
"y∈PYxiEyWy",
"xi",
",",
"where",
"PY",
"xi⊂V",
"ocY",
",",
"V",
"ocYis",
"the",
"entries",
"of",
"the",
"other",
"lan-",
"guage",
"in",
"the",
"multilingual",
"vocabulary",
",",
"0≤Wy",
"xi≤",
"1is",
"the",
"weight",
"of",
"the",
"candidate",
"y∈PY",
"xiand",
"/",
"summationtext",
"y∈PYxiWy",
"xi=",
"1",
"Given",
"the",
"model",
"Net",
",",
"we",
"have",
"4",
"steps",
"to",
"compute",
"[",
"C",
"]",
"xidynamically",
":",
"•Step",
"1",
":",
"We",
"set",
"Net",
"tothe",
"inference",
"mode",
"˜Net",
",",
"input",
"Ex+ELyto˜Net",
",",
"and",
"obtain",
"the",
"representation",
"hxi",
"&",
"Ly∈˜Net",
"(",
"Ex+ELy",
")",
"for",
"the",
"selected",
"token",
"xi",
"•Step",
"2",
":",
"We",
"factorize",
"Oyfrom",
"the",
"output",
"matrix",
"Oand",
"calculate",
"a",
"full",
"-",
"sized",
"set",
"Q=",
"(",
"hT",
"xi",
"&",
"LyOy0",
",",
"...",
",",
"hT",
"xi",
"&",
"LyOyv",
")",
",",
"where",
"vequals",
"the",
"size",
"of",
"V",
"ocY",
"•Step",
"3",
":",
"We",
"select",
"a",
"candidate",
"set",
"PY",
"xi=",
"(",
"Eyj",
",",
"...",
",",
"Eyk",
")",
"from",
"the",
"embedding",
"space",
",",
"ac-",
"cording",
"to",
"the",
"Top",
"-",
"K",
"dot",
"products",
"in",
"Q",
"•Step",
"4",
":",
"We",
"compute",
"the",
"weight",
"set",
"Wy",
"xi=",
"softmax",
"(",
"ET",
"yjEx",
",",
"...",
",",
"ET",
"ykEx",
")",
"and",
"the",
"final",
"output",
"E",
"[",
"C",
"]",
"xi=",
"/",
"summationtext",
"y∈PYxiEyWy",
"xi",
"Note",
"that",
",",
"multilingual",
"models",
"like",
"XLM",
"-",
"R",
"(",
"Con-",
"neau",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"I-MethodName",
"I-MethodName",
"O",
"O",
"O",
"O",
"O"
] |
[
",",
"2020",
")",
"do",
"not",
"require",
"language",
"embed-",
"dings",
",",
"i.e.",
",",
"eliminating",
"ELx",
"In",
"this",
"scenario",
",",
"we",
"can",
"simply",
"eliminate",
"ELyinStep",
"1",
"without",
"other",
"modifications",
",",
"and",
"we",
"still",
"obtain",
"cross",
"-",
"lingual",
"can-",
"didates",
"over",
"V",
"ocYinStep",
"2",
"to",
"compute",
"the",
"cross-",
"lingual",
"prototype",
"for",
"the",
"selected",
"token",
"xi",
"To",
"select",
"tokens",
"for",
"V",
"ocY",
",",
"the",
"minimum",
"fre-",
"quency",
"is",
"1e−5",
"in",
"the",
"monolingual",
"corpora",
"of",
"Lang",
"y",
"Meanwhile",
",",
"some",
"tokens",
"are",
"shared",
"among",
"3Empirical",
"studies",
"and",
"alternatives",
"of",
"Ex+ELxandEx",
"’s",
"nearest",
"neighbors",
"are",
"presented",
"in",
"Appendix",
"C.1.857",
"different",
"languages",
"We",
"set",
"the",
"minimum",
"frequency",
"of",
"shared",
"tokens",
"to",
"1e−3",
"in",
"the",
"monolingual",
"cor-",
"pora",
"These",
"settings",
"are",
"used",
"to",
"limit",
"the",
"searching",
"bound",
"for",
"more",
"meaningful",
"candidates",
"Inspiration",
"Our",
"recipe",
"takes",
"inspiration",
"from",
"early",
"experiments",
"We",
"pre",
"-",
"train",
"a",
"small",
"multi-",
"lingual",
"model",
"(",
"12",
"layers",
"and",
"256",
"d",
")",
"and",
"use",
"our",
"recipe",
"to",
"search",
"for",
"candidates",
"As",
"presented",
"in",
"Table",
"2",
",",
"a",
"multilingual",
"model",
"can",
"infer",
"some",
"cross-",
"lingual",
"candidates",
"with",
"our",
"recipe",
"because",
"of",
"the",
"cross",
"-",
"lingual",
"transfer",
"phenomenon",
",",
"and",
"we",
"can",
"generalize",
"these",
"candidates",
"for",
"cross",
"-",
"lingual",
"pro-",
"totypes",
"Meanwhile",
",",
"we",
"are",
"aware",
"that",
"the",
"mul-",
"tilingual",
"model",
"has",
"to",
"be",
"pre",
"-",
"trained",
"or",
"properly",
"initialized",
"in",
"order",
"to",
"infer",
"cross",
"-",
"lingual",
"candidates",
"by",
"itself",
"We",
"will",
"discuss",
"initialization",
"later",
"2.5",
"Alternation",
"between",
"[",
"M",
"]",
"and",
"[",
"C",
"]",
"x",
"In",
"our",
"experiment",
"(",
"see",
"row",
"12∼15of",
"Table",
"7",
"in",
"Appendix",
")",
",",
"we",
"find",
"that",
"we",
"can",
"get",
"bene-",
"fits",
"from",
"alternating",
"between",
"[",
"M",
"]",
"and",
"[",
"C",
"]",
"x",
"In-",
"tuitively",
",",
"only",
"using",
"[",
"C",
"]",
"xmight",
"perturb",
"bidirec-",
"tional",
"knowledge",
"and",
"result",
"in",
"the",
"lack",
"of",
"lan-",
"guage",
"knowledge",
",",
"whereas",
"the",
"model",
"can",
"learn",
"bidirectional",
"information",
"from",
"using",
"[",
"M",
"]",
"in",
"mul-",
"tilingual",
"MLM",
"pre",
"-",
"training",
"We",
"also",
"note",
"sim-",
"ilar",
"observations",
"in",
"previous",
"works",
"(",
"Chaudhary",
"et",
"al",
",",
"2020",
";",
"Ren",
"et",
"al",
",",
"2019a",
")",
",",
"which",
"use",
"trans-",
"lation",
"tables",
"for",
"pre",
"-",
"training",
"Another",
"side",
"ef-",
"fect",
"we",
"observe",
"is",
"that",
"the",
"model",
"might",
"pay",
"more",
"attention",
"to",
"\"",
"prototype",
"-",
"word",
"\"",
"translation",
"knowl-",
"edge",
"instead",
"of",
"understanding",
"bidirectional",
"knowl-",
"edge",
"Thus",
",",
"to",
"encourage",
"the",
"model",
"to",
"learn",
"both",
"strong",
"bidirectional",
"knowledge",
"from",
"[",
"M",
"]",
"and",
"cross",
"-",
"lingual",
"knowledge",
"from",
"[",
"C",
"]",
"x",
",",
"int",
"%",
"of",
"the",
"time",
"of",
"the",
"MLM",
"pre",
"-",
"training",
"time",
",",
"we",
"use",
"[",
"C",
"]",
"xfor",
"masking",
"For",
"the",
"remaining",
"(",
"100−t",
")",
"%",
"of",
"the",
"time",
",",
"we",
"still",
"use",
"[",
"M",
"]",
"Hence",
",",
"we",
"have",
"dual",
"objectives",
"in",
"multilingual",
"MLM",
"pre",
"-",
"training",
":",
"LMLM",
"=",
"L",
"[",
"C",
"]",
"x+L",
"[",
"M",
"]",
"With",
"these",
"dual",
"objectives",
"in",
"mind",
",",
"we",
"can",
"simply",
"extend",
"the",
"MLM",
"’s",
"masking",
"strategy",
"to",
":",
"(",
"[",
"SAME",
"]",
",",
"[",
"RAN",
"]",
",",
"[",
"M",
"]",
",",
"[",
"C",
"]",
"x",
")",
"with",
"(",
"10",
"%",
",",
"10",
"%",
",",
"(",
"80−t",
")",
"%",
",",
"t",
"%",
")",
"2.6",
"Discussion",
"We",
"discuss",
"some",
"important",
"components",
"of",
"our",
"method",
"For",
"these",
"discussions",
",",
"we",
"provide",
"em-",
"pirical",
"studies",
"and",
"show",
"the",
"observation",
"of",
"these",
"components",
"in",
"§",
"Robustness",
"and",
"Model",
"Variation"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricName",
"O",
"B-MetricName",
"I-MetricName"
] |
[
"[",
"M",
"]",
"vs.",
"[",
"C",
"]",
"x1",
")",
"[",
"M",
"]",
"is",
"static",
"in",
"the",
"embedding",
"space",
"with",
"an",
"explicit",
"entry",
",",
"used",
"by",
"running",
"alookup",
"operation",
"Meanwhile",
",",
"it",
"is",
"used",
"to",
"replace",
"all",
"randomly",
"selected",
"tokens",
",",
"which",
"is",
"unified",
".2",
")",
"In",
"contrast",
",",
"[",
"C",
"]",
"xiorE",
"[",
"C",
"]",
"xiis",
"dynamically",
"approxi-",
"mated",
"during",
"training",
",",
"which",
"is",
"token",
"-",
"wise",
"Choice",
"of",
"K",
"The",
"memory",
"usage",
"is",
"proportional",
"to",
"the",
"size",
"of",
"K",
"Meanwhile",
",",
"large",
"Kpotential",
"increases",
"noise",
"for",
"unambiguous",
"[",
"C",
"]",
"x.2",
")",
"On",
"the",
"other",
"hand",
",",
"a",
"small",
"Kmay",
"reduce",
"the",
"searching",
"bound",
"that",
"computing",
"proper",
"[",
"C",
"]",
"xis",
"hard"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"For",
"instance",
",",
"K=",
"1only",
"yields",
"median",
"improvements",
"in",
"our",
"experiment"
] | [
"O",
"O",
"O",
"B-HyperparameterName",
"B-HyperparameterValue",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Our",
"empirical",
"study",
"shows",
"that",
"it",
"is",
"robust",
"to",
"a",
"range",
"of",
"Kfrom",
"2",
"to",
"5",
",",
"considering",
"a",
"trade",
"-",
"off",
"between",
"GPU",
"memory",
"problems",
"and",
"expected",
"performance",
"improvements"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"B-HyperparameterValue",
"I-HyperparameterValue",
"I-HyperparameterValue",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Initialization",
"The",
"random",
"initialization",
"may",
"raise",
"problems",
"1",
")",
"xmay",
"find",
"some",
"geometric",
"close",
"but",
"irrelevant",
"tokens",
"with",
"large",
"dot",
"products",
"inV",
"ocY",
",",
"which",
"results",
"in",
"a",
"trivial",
"candidate",
"set",
"2",
")",
"Theinference",
"mode",
"with",
"random",
"initialization",
"is",
"trivial",
"To",
"this",
"end",
",",
"we",
"only",
"pre",
"-",
"train",
"the",
"multilin-",
"gual",
"model",
"by",
"MLM",
"with",
"[",
"M",
"]",
"at",
"the",
"first",
"several",
"iterations",
"for",
"warm",
"-",
"up",
"to",
"form",
"the",
"multilingual",
"em-",
"bedding",
"space",
"and",
"activate",
"the",
"inference",
"mode",
",",
"as",
"discussed",
"in",
"§",
"Inspiration",
"After",
"the",
"warm",
"-",
"up",
",",
"the",
"multilingual",
"embedding",
"space",
"and",
"the",
"inference",
"mode",
"are",
"initialized",
"in",
"a",
"few",
"-",
"shot",
"style",
"somewhat",
"to",
"avoid",
"trivial",
"candidates",
"Then",
",",
"we",
"run",
"the",
"al-",
"ternation",
"In",
"our",
"experiments",
",",
"we",
"find",
"that",
"this",
"warm",
"-",
"up",
"can",
"help",
"the",
"model",
"obtain",
"new",
"samples",
"with",
"cross",
"-",
"lingual",
"prototypes",
"from",
"the",
"other",
"lan-",
"guage",
"Efficiency",
"On",
"-",
"the",
"-",
"fly",
"[",
"C",
"]",
"xwill",
"increase",
"the",
"train-",
"ing",
"time",
"However",
",",
"only",
"a",
"subset",
"of",
"tokens",
"(",
"typi-",
"cally",
",",
"15",
"%",
"(",
"Devlin",
"et",
"al",
",",
"2019",
")",
")",
"of",
"the",
"input",
"text",
"stream",
"is",
"selected",
"for",
"masking",
",",
"and",
"we",
"only",
"need",
"to",
"compute",
"[",
"C",
"]",
"xfor",
"a",
"sub",
"-",
"set",
"of",
"all",
"the",
"selected",
"tokens",
"In",
"our",
"experiment",
",",
"we",
"find",
"our",
"method",
"spends",
"additional",
"≈15",
"%",
"time",
"on",
"training",
"Tokenization",
"Tokenizations",
"generating",
"“",
"middle",
"\"",
"tokens",
",",
"sub",
"-",
"tokens",
",",
"or",
"non",
"-",
"standard",
"word",
"tokens",
"might",
"impact",
"[",
"C",
"]",
"x",
",",
"e.g.",
",",
"BPE",
"However",
",",
"the",
"impact",
"is",
"relatively",
"small",
"given",
"that",
":",
"1",
")",
":",
"the",
"vocabularies",
"and",
"monolingual",
"corpora",
"are",
"dominant",
"by",
"the",
"stan-",
"dard",
"words",
"rather",
"than",
"non",
"-",
"standard",
"word",
"token",
",",
"e.g.",
",",
"over",
"50",
"%",
"BPE",
"vocabulary",
"for",
"translation",
"task",
"De↔En",
")",
"are",
"standard",
"words",
"and",
"they",
"make",
"up",
"for",
"over",
"the",
"80",
"%",
"of",
"the",
"total",
"token",
"frequency",
"in",
"the",
"monolingual",
"corpora",
";",
"2",
")",
":",
"all",
"the",
"representations",
"are",
"contextualized",
"that",
"sub",
"-",
"tokens",
"and",
"non",
"-",
"standard858",
"400k",
"step",
"training",
"50k",
"step",
"training",
"#",
"1",
"It",
"was",
"hampered",
"by",
"the",
"need",
"for",
"ranges",
"to",
"be",
"estimated",
"by",
"eye",
",",
"which",
"introduced",
"significant",
"in",
"@",
"@",
"accuracy"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName"
] |
[
"[",
"EOS",
"]",
"Reference",
"Erschwert",
"wurde",
"dies",
"durch",
"die",
"Notwendigkeit",
",",
"Entfernungen",
"mit",
"dem",
"Auge",
"abzuschätzen",
",",
"was",
"zu",
"erheblichen",
"Ungenauigkeiten",
"führte",
"[",
"EOS",
"]",
"Masked",
"It",
"[",
"C",
"]",
"x1hampered",
"[",
"C",
"]",
"x3",
"[",
"C",
"]",
"x4need",
"[",
"C",
"]",
"x6ranges",
"to",
"[",
"C",
"]",
"x9estimated",
"by",
"eye",
",",
"which",
"[",
"C",
"]",
"x15significant",
"[",
"C",
"]",
"x17",
"[",
"C",
"]",
"x18",
"[",
"EOS",
"]",
"was",
"=",
"[",
"C",
"]",
"x1war",
",",
"wurde",
",",
"brach",
",",
"und",
",",
"als",
"by",
"=",
"[",
"C",
"]",
"x3in",
",",
",",
",",
"durch",
"in",
",",
"von",
",",
",",
"the",
"=",
"[",
"C",
"]",
"x4den",
",",
"die",
",",
"der",
",",
"den",
",",
"einem",
"for",
"=",
"[",
"C",
"]",
"x6für",
",",
"in",
",",
"dafür",
"in",
",",
",",
"und",
"be",
"=",
"[",
"C",
"]",
"x9des",
",",
",",
",",
"ben",
"stellt",
",",
"Bau",
",",
"einem",
"introduced",
"=",
"[",
"C",
"]",
"x15lehnte",
",",
"Schwei",
"@",
"@",
",",
"löste",
"in",
",",
"/",
",",
"von",
"in",
"@",
"@",
"=",
"[",
"C",
"]",
"x17",
"in",
",",
"Gebäude",
"@",
"@",
",",
"@",
"-",
"@",
"in",
",",
"(",
",",
"@",
"-",
"@",
"accuracy",
"=",
"[",
"C",
"]",
"x18Seh",
"@",
"@",
",",
"ographie",
",",
"Bewertung",
"geber",
",",
"er",
",",
"studium",
"#",
"2",
"Sie",
"befindet",
"sich",
"auf",
"425",
"Meter",
"Höhe",
"nahe",
"dem",
"Schlos",
"@",
"@",
"sberg",
"[",
"EOS",
"]",
"Reference",
"It",
"is",
"located",
"at",
"an",
"altitude",
"of",
"425",
"meters",
"near",
"the",
"Schlossberg",
"[",
"EOS",
"]",
"Masked",
"[",
"C",
"]",
"x0",
"[",
"C",
"]",
"x1sich",
"auf",
"425",
"[",
"C",
"]",
"x5",
"[",
"C",
"]",
"x6",
"[",
"C",
"]",
"x7dem",
"Schlos",
"@",
"@",
"sberg",
"[",
"EOS",
"]",
"Sie",
"=",
"[",
"C",
"]",
"x0It",
",",
"She",
",",
"He",
"leaves",
",",
"breaks",
",",
"Geography",
"auf",
"=",
"[",
"C",
"]",
"x4in",
",",
"at",
",",
"on",
"the",
",",
"a",
",",
"29",
"@",
"@",
"Meter",
"=",
"[",
"C",
"]",
"x5metres",
",",
"meters",
",",
"feet",
"@",
"-",
"@",
",",
",",
",",
"in",
"Höhe",
"=",
"[",
"C",
"]",
"x6altitude",
",",
"height",
",",
"elevation",
",",
",",
",",
"in",
",",
"an",
"nahe",
"=",
"[",
"C",
"]",
"x7near",
",",
"Near",
",",
"close",
"in",
",",
"an",
",",
",",
"Table",
"2",
":",
"Inspiration",
"of",
"[",
"C",
"]",
"xfrom",
"multilingual",
"training",
"References",
"are",
"obtained",
"from",
"Google",
"Translation",
"We",
"use",
"a",
"pre",
"-",
"trained",
"small",
"XLM",
"model",
"on",
"{",
"En",
",",
"De",
"}",
"To",
"obtain",
"more",
"examples",
",",
"we",
"randomly",
"compute",
"[",
"C",
"]",
"xfor",
"40",
"%",
"of",
"tokens.",
"@",
"@",
"is",
"the",
"continuing",
"subword",
"prefix",
"bold",
"denotes",
"a",
"strong",
"candidate",
"that",
"is",
"a",
"parallel",
",",
"analogical",
",",
"or",
"relevant",
"token",
"/",
"word",
"(",
"or",
"its",
"variation",
")",
"in",
"other",
"languages",
"Our",
"method",
"can",
"cover",
"multiple",
"morphological",
"or",
"relevant",
"candidates",
"(",
"e.g.",
",",
"<",
"den",
",",
"die",
",",
"der",
">",
"in",
"#",
"1",
"[",
"C",
"]",
"x4",
")",
"for",
"generalizing",
"information",
"by",
"weighted",
"average",
"word",
"tokens",
"still",
"represent",
"semantics",
"and",
"syntactic",
"meanings",
"related",
"to",
"their",
"original",
"standard",
"words",
"(",
"refer",
"to",
"the",
"case",
"study",
"in",
"Appendix",
"C.2",
")",
"3",
"Empirical",
"Study",
"and",
"Experiment",
"All",
"the",
"links",
"of",
"datasets",
",",
"libraries",
",",
"scripts",
",",
"and",
"tools",
"marked",
"with",
"⋄are",
"listed",
"in",
"Appendix",
"F",
"A",
"preview",
"version",
"of",
"the",
"code",
"is",
"submitted",
",",
"and",
"we",
"will",
"open",
"the",
"source",
"code",
"on",
"GitHub",
"Pre",
"-",
"training",
"Setting",
"We",
"use",
"Adam",
"optimizer",
"(",
"Kingma",
"and",
"Ba",
",",
"2015",
")",
"with",
"β1=",
"0.9",
",",
"β2=",
"0.999",
",",
"ϵ=",
"1e−8",
",",
"warm",
"_",
"upstep",
"(",
"Vaswani",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"B-HyperparameterValue",
"O",
"B-HyperparameterName",
"B-HyperparameterValue",
"O",
"B-HyperparameterName",
"B-HyperparameterValue",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"O",
"O",
"O",
"O"
] |
[
",",
"2017",
")",
"andlr=",
"1e−4"
] | [
"O",
"O",
"O",
"B-HyperparameterName",
"B-HyperparameterValue"
] |
[
"Dropout",
"regularization",
"is",
"set",
"to",
"rate",
"=",
"0.1"
] | [
"B-HyperparameterName",
"I-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterValue"
] |
[
"Readers",
"can",
"refer",
"to",
"Appendix",
"D.1",
"for",
"details",
"Model",
"Configuration",
"Our",
"Transformer",
"model",
"(",
"Vaswani",
"et",
"al",
",",
"2017",
")",
"is",
"identical",
"to",
"XLM",
"(",
"Lample",
"and",
"Conneau",
",",
"2019",
")",
",",
"which",
"consists",
"of",
"a",
"6",
"-",
"layer",
"en-",
"coder",
"and",
"6",
"-",
"layer",
"decoder",
"with",
"1024",
"word",
"embed-",
"ding",
"size",
"and",
"hidden",
"size",
"and",
"4096",
"feed",
"-",
"forward",
"filter",
"size"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"We",
"add",
"a",
"learnable",
"language",
"embedding",
"and",
"a",
"learnable",
"position",
"embedding",
"to",
"each",
"token",
"of",
"the",
"input",
"sentence",
"for",
"the",
"encoder",
"and",
"decoder",
"(",
"PandLin",
"Eq.1",
")",
"We",
"have",
"some",
"default",
"configu-",
"rations",
"for",
"our",
"method",
"based",
"on",
"the",
"study",
"of",
"model",
"robustness",
"(",
"see",
"§",
"Robustness",
"and",
"Model",
"Variation",
")",
":",
"1",
")",
"t",
"%",
"=",
"40",
"%",
"that",
"we",
"make",
"a",
"balance",
"between",
"the",
"two",
"objectives",
":",
"[",
"M",
"]",
"and",
"[",
"C",
"]",
"x",
";",
"2",
")",
"K=",
"3",
"that",
"we",
"consider",
"top-3",
"candidates",
"for",
"the",
"cross",
"-",
"lingual",
"prototypes",
";",
"3",
")",
"the",
"warm",
"-",
"up",
"step",
"is",
"50k",
"that",
"[",
"M",
"]",
"is",
"only",
"used",
"at",
"the",
"first",
"50k",
"iterations",
";",
"4",
")",
"we",
"consider",
"BPE",
"for",
"tokenization",
"in",
"all",
"our",
"experiments"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"O",
"B-HyperparameterValue",
"I-HyperparameterValue",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"O",
"B-HyperparameterValue",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Multilingual",
"Task",
"We",
"consider",
"three",
"multilin-",
"gual",
"tasks",
":",
"1",
")",
"UNMT",
"for",
"evaluation",
"on",
"translation",
"tasks",
",",
"2",
")",
"cross",
"-",
"lingual",
"word",
"similarity",
"for",
"evalua-",
"tion",
"on",
"cross",
"-",
"lingual",
"embedding",
"tasks",
",",
"and",
"3",
")",
"zero-",
"shot",
"cross",
"-",
"lingual",
"classification",
"for",
"evaluation",
"on",
"cross",
"-",
"lingual",
"transfer",
"tasks"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"3.1",
"MLM",
"Instance",
"We",
"adapt",
"our",
"method",
"to",
"three",
"MLM",
"instances",
"to",
"pre",
"-",
"train",
"the",
"multilingual",
"model",
":",
"1",
")",
"XLM",
"(",
"Lample",
"and",
"Conneau",
",",
"2019",
")",
",",
"2",
")",
"MASS",
"(",
"Song",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O"
] |
[
",",
"2019",
")",
",",
"and3",
")",
"mBART",
"(",
"Liu",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O"
] |
[
",",
"2020",
")",
",",
"which",
"can",
"be",
"used",
"to",
"pre",
"-",
"train",
"a",
"multilingual",
"model",
"Readers",
"can",
"refer",
"to",
"the",
"original",
"report",
"or",
"Appendix",
"D.2",
"for",
"more",
"instructions",
"on",
"these",
"MLM",
"instances",
"Sig-",
"nificantly",
",",
"to",
"minimize",
"changes",
"for",
"evaluation",
"and",
"comparison",
",",
"we",
"only",
"have",
"two",
"changes",
"The",
"first",
"change",
"we",
"make",
"is",
"extending",
"the",
"masking",
"strategy",
":",
"(",
"[",
"SAME",
"]",
",",
"[",
"RAN",
"]",
",",
"[",
"M",
"]",
")",
"with",
"(",
"10",
"%",
",",
"10",
"%",
",80",
"%",
")",
"to",
"(",
"[",
"SAME",
"]",
",",
"[",
"RAN",
"]",
",",
"[",
"M",
"]",
",",
"[",
"C",
"]",
"x",
")",
"with",
"(",
"10",
"%",
",",
"10",
"%",
",",
"(",
"80−t",
")",
"%",
",",
"t",
"%",
")",
"Secondly",
",",
"as",
"presented",
"in",
"Table",
"1",
",",
"we",
"only",
"apply",
"CLPM",
"to",
"the",
"input",
"of",
"the",
"source",
"side",
"or",
"the",
"encoder",
"and",
"do",
"not",
"change",
"the",
"shifted",
"input",
"of",
"the",
"decoder",
"in",
"these",
"MLM",
"instances"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Any",
"other",
"component",
"is",
"identical",
"to",
"the",
"reported",
"MLM",
"instances",
"We",
"reimplement",
"all",
"the",
"baseline",
"models",
"on",
"our",
"machine",
"with",
"our",
"configurations",
",",
"using",
"official",
"XLM⋄",
",",
"Tensor2Tensor",
"⋄",
",",
"and",
"HuggingFace",
"⋄as",
"ref-",
"erences",
"We",
"compare",
"the",
"results",
"of",
"our",
"reimple-",
"mentation",
"with",
"the",
"reported",
"results",
"on",
"the",
"same",
"test",
"set",
"to",
"ensure",
"that",
"the",
"difference",
"is",
"less",
"than",
"2",
"%",
"in",
"overall",
"performance",
"(",
"see",
"Appendix",
"E",
"for",
"result859",
"comparison",
")",
"Then",
",",
"we",
"can",
"confirm",
"our",
"reimple-",
"mentation",
"3.2",
"UNMT",
"Setup",
"We",
"consider",
"similar",
"language",
"pairs",
"{",
"De",
",",
"Ro",
"}",
"↔En",
",",
"using",
"the",
"same",
"dataset",
"and",
"test",
"set",
"as",
"previous",
"works",
"(",
"Lample",
"and",
"Conneau",
",",
"2019",
")"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Meanwhile",
",",
"we",
"share",
"the",
"FLoRes",
"⋄",
"(",
"Guzmán",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O"
] |
[
",",
"2019",
")",
"task",
"to",
"evaluate",
"a",
"dissimilar",
"language",
"pair",
"Ne↔English",
"(",
"Nepali",
")",
"We",
"learn",
"shared",
"BPE",
"(",
"Sennrich",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-DatasetName",
"I-DatasetName",
"O",
"O",
"O",
"O"
] |
[
",",
"2016b",
")",
",",
"selecting",
"the",
"most",
"fre-",
"quent",
"60",
"K",
"codes",
"from",
"paired",
"languages",
"with",
"the",
"same",
"criteria",
"in",
"Lample",
"and",
"Conneau",
"(",
"2019",
")",
"The",
"model",
"is",
"pre",
"-",
"trained",
"around",
"400",
"K",
"iterations",
"on",
"only",
"monolingual",
"corpora",
"in",
"different",
"languages"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterValue",
"I-HyperparameterValue",
"B-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"And",
",",
"after",
"around",
"400",
"K",
"training",
"iterations",
"for",
"trans-",
"lation",
"with",
"the",
"standard",
"pipeline",
"⋄",
"(",
"Artetxe",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"B-HyperparameterValue",
"I-HyperparameterValue",
"B-HyperparameterName",
"I-HyperparameterName",
"O",
"B-TaskName",
"I-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
",",
"2018b",
";",
"Song",
"et",
"al",
",",
"2019",
")",
",",
"according",
"to",
"baseline",
"models",
"’",
"BLEU",
"scripts",
",",
"we",
"report",
"BLEU",
"computed",
"bymulti-BLEU.perl",
"⋄orsacreBleu",
"⋄",
"(",
"Post",
",",
"2018",
")",
"with",
"default",
"rules"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricName",
"O",
"O",
"O",
"O",
"B-MetricName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"See",
"more",
"details",
"in",
"Appendix",
"D.3",
"Result",
"Table",
"3",
"shows",
"the",
"results",
"on",
"the",
"{",
"De",
",",
"Ro",
",",
"Ne",
"}",
"↔",
"Entest",
"sets",
"Applying",
"[",
"C",
"]",
"x",
"consistently",
"improves",
"the",
"performance",
"of",
"base-",
"line",
"models",
"on",
"all",
"the",
"similar",
"language",
"pairs",
"by",
"3",
"%",
"∼8",
"%",
"and",
"on",
"the",
"dissimilar",
"pair",
"by",
"2.5∼7",
"BLEU"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"I-MetricValue",
"I-MetricValue",
"I-MetricValue",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"B-MetricName"
] |
[
"The",
"performance",
"on",
"the",
"dissimilar",
"pair",
"is",
"very",
"close",
"to",
"SOTA",
":",
"mBART25",
"(",
"Liu",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O"
] |
[
",",
"2020",
")",
",",
"but",
"they",
"use",
"25",
"languages",
"from",
"CC25",
"(",
"Wenzek",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-DatasetName",
"O",
"O",
"O",
"O"
] |
[
",",
"2020",
")",
"for",
"pre",
"-",
"training",
"Our",
"method",
"slightly",
"outperforms",
"two",
"dictionary",
"-",
"based",
"works",
"(",
"Dufter",
"and",
"Schütze",
",",
"2020",
";",
"Chaudhary",
"et",
"al",
",",
"2020",
")",
"which",
"require",
"static",
"translation",
"tables",
"from",
"pre",
"-",
"trained",
"word",
"models",
",",
"golden",
"dictionaries",
",",
"or",
"bilingual",
"lex-",
"icon",
"induction",
"(",
"e.g.",
",",
"UBWE",
")",
"Intuitively",
",",
"as",
"re-",
"ported",
"in",
"(",
"Artetxe",
"et",
"al",
",",
"2020",
";",
"Kementchedjhieva",
"et",
"al",
",",
"2019",
";",
"Czarnowska",
"et",
"al",
",",
"2019",
";",
"Vania",
"and",
"Lopez",
",",
"2017",
")",
",",
"such",
"word",
"translation",
"tables",
"are",
"re-",
"ported",
"to",
"misrepresent",
"morphological",
"variations",
"and",
"are",
"not",
"contextualized",
"properly",
",",
"which",
"limit",
"the",
"improvements",
"for",
"sentence",
"translation",
"For",
"further",
"analyses",
",",
"we",
"conduct",
"a",
"case",
"study",
"to",
"observe",
"the",
"attention",
"weights",
"on",
"[",
"C",
"]",
"xafter",
"pre-",
"training",
",",
"which",
"is",
"visualized",
"in",
"Appendix",
"C.2",
"We",
"observe",
"that",
"the",
"model",
"outputs",
"prominent",
"attention",
"weights",
"on",
"[",
"C",
"]",
"xfor",
"predicting",
"replaced",
"tokens",
",",
"so",
"that",
"it",
"relies",
"on",
"[",
"C",
"]",
"x",
"In",
"other",
"words",
",",
"the",
"model",
"understands",
"[",
"C",
"]",
"xin",
"the",
"context",
"We",
"can",
"confirm",
"the",
"effectiveness",
"Concretely",
",",
"CLPM",
"shows",
"sig-",
"nificant",
"effectiveness",
"on",
"nouns",
",",
"entities",
",",
"terminol-",
"Figure",
"1",
":",
"Discriminator",
"performance"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"The",
"discriminator",
"is",
"trained",
"to",
"recognize",
"which",
"language",
"an",
"embedding",
"or",
"a",
"representation",
"belongs",
"to",
"and",
"makes",
"zero",
"-",
"shot",
"clas-",
"sification",
"for",
"a",
"prototype"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O"
] |
[
"We",
"use",
"all",
"the",
"embedding",
"instances",
"to",
"train",
"the",
"discriminator",
"This",
"figure",
"indicates",
"that",
"CLMP",
"introduces",
"unseen",
"cross",
"-",
"lingual",
"prototypes",
"for",
"the",
"model",
"instead",
"of",
"embedding",
"instances"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"ogy",
"words",
",",
"etc",
",",
"where",
"the",
"attention",
"weights",
"on",
"the",
"corresponding",
"[",
"C",
"]",
"xare",
"dominant",
"Meanwhile",
",",
"the",
"model",
"can",
"understand",
"phrases",
",",
"sub",
"-",
"tokens",
",",
"and",
"syntax",
"structures",
"to",
"predict",
"a",
"replaced",
"token",
"of",
"the",
"phrase",
"because",
"the",
"model",
"pays",
"equal",
"/",
"similar",
"at-",
"tention",
"to",
"each",
"token",
"of",
"the",
"phrase",
"We",
"attribute",
"this",
"phenomenon",
"to",
"both",
"the",
"alternation",
"between",
"[",
"C",
"]",
"xand",
"[",
"M",
"]",
"and",
"involving",
"neighboring",
"tokens",
"in",
"{",
"[",
"C",
"]",
"xi",
",",
"xj\\i",
"}",
"→",
"xithat",
"the",
"model",
"captures",
"to-",
"ken",
"dependencies",
"from",
"the",
"cross",
"-",
"lingual",
"prototype",
"or",
"a",
"synonym",
"in",
"the",
"other",
"language",
"Finally",
",",
"the",
"employment",
"of",
"multiple",
"candidates",
"is",
"important",
"because",
"the",
"model",
"could",
"learn",
"morphological",
"or",
"relevant",
"variations",
"from",
"[",
"C",
"]",
"xin",
"the",
"other",
"language",
"(",
"refer",
"to",
"Appendix",
"C.1",
")",
",",
"e.g.",
",",
"understanding",
"rele-",
"vant",
"variations",
"<",
"welches",
",",
"welcher",
",",
"welche",
">",
"from",
"[",
"C",
"]",
"x",
",",
"which",
"is",
"essential",
"for",
"further",
"translation",
"learn-",
"ing",
"in",
"unsupervised",
"manners",
"Dose",
"CLMP",
"introduce",
"new",
"samples",
"with",
"cross-",
"lingual",
"prototypes",
"from",
"the",
"other",
"language",
"?"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"In",
"addition",
"to",
"§",
"Case",
"Study",
",",
"we",
"are",
"still",
"interested",
"in",
"the",
"representation",
"of",
"E",
"[",
"C",
"]",
"xor",
"whether",
"CLMP",
"intro-",
"duces",
"new",
"examples",
"with",
"cross",
"-",
"lingual",
"prototypes",
"from",
"the",
"other",
"language",
"Intuitively",
",",
"if",
"the",
"weights",
"obtained",
"in",
"Step",
"4",
"are",
"{",
"c1=",
"0.9",
",",
"c2=",
"0.05",
",",
"c3=",
"0.05",
"}",
",",
"the",
"representation",
"is",
"similar",
"to",
"the",
"candi-",
"datec1",
",",
"and",
"then",
"c1is",
"a",
"soft",
"translation",
"of",
"x",
"If",
"the",
"weights",
"are",
"{",
"c1=",
"0.4",
",",
"c2=",
"0.3",
",",
"c3=",
"0.3",
"}",
",",
"the",
"representation",
"could",
"be",
"different",
"from",
"any",
"one",
"of",
"{",
"c1",
",",
"c2",
",",
"c3",
"}",
"Thus",
",",
"the",
"representation",
"depends",
"on",
"the",
"contributions",
"of",
"the",
"candidates",
"To",
"further",
"understand",
"E",
"[",
"C",
"]",
"x",
",",
"we",
"jointly",
"train",
"a",
"discrimina-",
"tor",
"to",
"distinguish",
"between",
"two",
"languages",
"in",
"the",
"pre",
"-",
"training",
"phase",
"The",
"discriminator",
"is",
"trained860",
"Language",
"pair",
"De↔En",
"Ro",
"↔En",
"Ne",
"↔En",
"multi-BLEU.perl",
"⋄with",
"default",
"rules",
"XLM",
"(",
"Lample",
"et",
"al",
",",
"2018c",
")",
"34.3",
"26.4",
"31.8",
"33.3",
"0.5",
"0.1",
"+",
"word",
"translation",
"tables",
"(",
"Chaudhary",
"et",
"al",
",",
"2020",
")",
"⋆",
"35.1",
"27.4",
"33.6",
"34.4",
"4.1",
"2.2",
"+",
"[",
"C",
"]",
"x",
"35.9",
"28.1",
"34.4",
"35.3",
"6.6",
"2.8",
"MASS",
"(",
"Song",
"et",
"al",
",",
"2019",
")",
"35.2",
"28.3",
"33.1",
"35.2",
"+",
"nearest",
"neighbor",
"from",
"UBWE",
"(",
"Dufter",
"and",
"Schütze",
",",
"2020",
")",
"⋆",
"36.1",
"28.8",
"34.1",
"36.4",
"5.1",
"2.8",
"+",
"[",
"C",
"]",
"x",
"36.7",
"29.2",
"34.7",
"36.9",
"7.1",
"3.4",
"sacreBleu",
"⋄with",
"standard",
"settings",
":",
"nrefs:1|case",
":",
"mixed|eff",
":",
"no|tok:13a|smooth",
":",
"exp|version:2.0.0",
"mBART",
"(",
"Liu",
"et",
"al",
",",
"2020",
")",
"+",
"CC25",
"(",
"Wenzek",
"et",
"al",
",",
"2020",
")",
"34.0",
"29.8",
"30.5",
"35.0",
"10.0",
"4.4",
"+",
"[",
"C",
"]",
"x",
"(",
"w",
"/",
"o",
"CC25",
")",
"35.4",
"30.1",
"32.5",
"36.7",
"7.0",
"3.2",
"Table",
"3",
":",
"Performance",
"of",
"UNMT",
"⋆are",
"reimplemented",
"UBWE",
"stands",
"for",
"unsupervised",
"bilingual",
"word",
"embedding",
"Translation",
"tables",
"or",
"UBWE",
"are",
"static",
"We",
"use",
"the",
"same",
"transformer",
"models",
",",
"BPE",
"size",
",",
"corpora",
",",
"tokenization",
",",
"and",
"BLEU",
"as",
"the",
"baseline",
"models",
"(",
"see",
"more",
"details",
"in",
"Appendix",
"D.3",
")",
"to",
"recognize",
"which",
"language",
"an",
"embedding",
"or",
"a",
"representation",
"belongs",
"to",
"We",
"use",
"all",
"the",
"embed-",
"ding",
"instances",
"to",
"train",
"the",
"discriminator",
"Then",
",",
"we",
"make",
"zero",
"-",
"shot",
"classification",
"for",
"E",
"[",
"C",
"]",
"xto",
"ob-",
"serve",
"which",
"language",
"E",
"[",
"C",
"]",
"xbelongs",
"to"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"We",
"re-",
"port",
"the",
"result",
"in",
"Figure",
"1",
"This",
"figure",
"suggests",
"that",
"CLMP",
"introduces",
"unseen",
"cross",
"-",
"lingual",
"pro-",
"totypes",
"for",
"the",
"model",
"We",
"suspect",
"that",
"ECxpo-",
"tentially",
"yields",
"a",
"generalized",
"representation",
"from",
"multiple",
"relevant",
"candidates",
"in",
"other",
"languages",
"This",
"is",
"different",
"from",
"the",
"method",
"family",
"based",
"on",
"translation",
"tables",
"Significantly",
",",
"translation",
"tables",
"are",
"instances",
"/",
"embeddings",
"in",
"the",
"embedding",
"space",
",",
"whereas",
"cross",
"-",
"lingual",
"prototypes",
"do",
"not",
"exist",
"in",
"the",
"embedding",
"spaces",
"and",
"are",
"new",
"generalized",
"samples",
"for",
"the",
"model",
"3.3",
"Robustness",
"and",
"Model",
"Variation",
"We",
"have",
"some",
"default",
"configurations",
",",
"as",
"presented",
"in",
"row",
"2",
"of",
"Table",
"4"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricName",
"I-MetricName",
"I-MetricName",
"I-MetricName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"This",
"combination",
"is",
"obtained",
"in",
"our",
"experiments",
"We",
"report",
"the",
"results",
"to",
"observe",
"the",
"impact",
"of",
"K",
"(",
"the",
"number",
"of",
"cross",
"-",
"lingual",
"can-",
"didates",
")",
",",
"the",
"warm",
"-",
"up",
"initialization",
",",
"the",
"tokeniza-",
"tion",
"method",
",",
"and",
"the",
"alternation",
"t",
"%",
"in",
"Appendix",
"B"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"O",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"O",
"O",
"O"
] |
[
"Meanwhile",
",",
"in",
"this",
"experiment",
",",
"we",
"discuss",
"a",
"mean",
"average",
"style",
"for",
"cross",
"-",
"lingual",
"candidates",
"instead",
"of",
"the",
"weighted",
"average",
"used",
"in",
"the",
"default",
"configura-",
"tion",
",",
"reporting",
"results",
"in",
"Appendix",
"B",
"Additionally",
",",
"we",
"study",
"alternatives",
"for",
"initialization",
"and",
"training",
"efficiency",
"The",
"result",
"is",
"presented",
"in",
"Table",
"7",
"For",
"consistency",
",",
"the",
"row",
"number",
"is",
"consistent",
"with",
"the",
"full",
"results",
"in",
"Appendix",
"B",
"Row",
"11",
"As",
"aforementioned",
",",
"CLPM",
"requires",
"ad-",
"ditional",
"time",
"to",
"compute",
"[",
"C",
"]",
"x"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"To",
"be",
"fair",
",",
"we",
"reduce",
"the",
"training",
"steps",
",",
"so",
"that",
"the",
"training",
"time",
"is",
"almost",
"similar",
"to",
"the",
"baseline",
"model",
"(",
"row",
"1",
")",
"CLPM",
"out-",
"performs",
"the",
"baseline",
"model",
"but",
"requires",
"fewer",
"training",
"steps",
",",
"which",
"indicates",
"that",
"the",
"explicit",
"and",
"principled",
"cross",
"-",
"lingual",
"forward",
"pass",
"is",
"more",
"ef",
"-",
"ficient",
"(",
"per",
"step",
")",
"than",
"implicit",
"isomorphic",
"space",
"formation",
"for",
"cross",
"-",
"linguality"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Row",
"17",
"We",
"use",
"UBWE",
"(",
"unsupervised",
"bilingual",
"word",
"embedding",
")",
"to",
"initialize",
"the",
"bilingual",
"em-",
"bedding",
"space",
"In",
"the",
"first",
"50k",
"pre",
"-",
"training",
"steps",
"(",
"equal",
"to",
"default",
"warm",
"-",
"up",
"steps",
")",
",",
"since",
"the",
"model",
"parameters",
"are",
"still",
"randomly",
"initialized",
",",
"we",
"do",
"not",
"follow",
"Step",
"1",
",",
"2",
",",
"and",
"3",
"in",
"on",
"-",
"the",
"-",
"fly",
"[",
"C",
"]",
"xand",
"directly",
"find",
"relevant",
"candidates",
"based",
"on",
"the",
"dot",
"products",
"ET",
"yiEx",
",",
"i.e.",
",",
"only",
"need",
"Step",
"4"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Intuitively",
",",
"ET",
"yiExis",
"reliable",
"to",
"rank",
"the",
"candidates",
"and",
"com-",
"pute",
"the",
"weights",
"for",
"[",
"C",
"]",
"xbecause",
"UBWE",
"provides",
"cross",
"-",
"lingual",
"entries",
"After",
"50k",
"pre",
"-",
"training",
"steps",
",",
"we",
"normally",
"run",
"on",
"-",
"the",
"-",
"fly",
"[",
"C",
"]",
"x"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"We",
"observe",
"that",
"adapting",
"UBWE",
"consistently",
"improves",
"the",
"perfor-",
"mance",
"by",
"2",
"%",
"on",
"the",
"similar",
"language",
"and",
"0.5∼1",
"BLEU",
"on",
"the",
"dissimilar",
"language",
"because",
"UBWE",
"provides",
"additional",
"cross",
"-",
"lingual",
"supervision"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"O",
"B-HyperparameterValue",
"I-HyperparameterValue",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterValue",
"B-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"See",
"all",
"the",
"results",
"in",
"Table",
"8",
"Row",
"18",
"Vuli´c",
"et",
"al",
"(",
"2020",
")",
"suggest",
"seed",
"dictionar-",
"ies",
"for",
"unsupervised",
"tasks",
"in",
"practice",
"Following",
"this",
"idea",
",",
"we",
"download",
"a",
"1k",
"seed",
"dictionary",
"from",
"Panlex",
"⋄",
"In",
"the",
"first",
"50k",
"pre",
"-",
"training",
"steps",
",",
"we",
"simply",
"replace",
"the",
"selected",
"token",
"with",
"its",
"trans-",
"lation",
"in",
"the",
"seed",
"dictionary",
"For",
"the",
"out",
"-",
"of",
"-",
"the-",
"dictionary",
"but",
"selected",
"token",
",",
"we",
"replace",
"it",
"with",
"normal",
"[",
"M",
"]",
"After",
"50k",
"pre",
"-",
"training",
"steps",
",",
"if",
"the",
"selected",
"token",
"is",
"in",
"the",
"dictionary",
",",
"the",
"translation",
"is",
"added",
"to",
"[",
"C",
"]",
"xas",
"a",
"candidate",
"in",
"Step",
"4",
"when",
"run-",
"ning",
"on",
"-",
"the",
"-",
"fly",
"[",
"C",
"]",
"x",
"We",
"find",
"that",
"compared",
"to",
"the",
"UBWE",
"scenario",
",",
"this",
"adaptation",
"achieves",
"similar",
"results",
"on",
"the",
"rich",
"-",
"resource",
"language",
"De↔En",
"(",
"+1.5",
"%",
")",
"but",
"stronger",
"results",
"on",
"the",
"dissimilar",
"lan-",
"guage",
"Ne↔En",
"(",
"+8",
"%",
")"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"I-MetricValue",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"I-MetricValue",
"I-MetricValue",
"I-MetricValue"
] |
[
"All",
"the",
"results",
"are",
"pre-",
"sented",
"in",
"Table",
"8.861",
"Row",
"Model",
"t",
"Tokenization",
"Warm",
"-",
"up",
"Steps",
"K",
"[",
"C",
"]",
"xtype",
"De↔En",
"1",
"[",
"M",
"]",
"(",
"baseline",
")",
"-",
"BPE",
"-",
"400",
"K",
"-",
"-",
"34.3",
"26.4",
"2",
"[",
"C",
"]",
"x",
"(",
"our",
"baseline",
",",
"default",
")",
"40",
"%",
"BPE",
"50",
"K",
"400",
"K",
"3",
"weighted",
"35.9",
"28.1",
"11",
"[",
"C",
"]",
"x",
"+",
"+",
"+",
"350",
"K",
"(",
"similar",
"training",
"time",
")",
"+",
"+",
"35.1",
"27.2",
"17",
"[",
"C",
"]",
"x",
"+",
"+",
"UBWE",
"+",
"+",
"+",
"36.5",
"28.8",
"18",
"[",
"C",
"]",
"x",
"+",
"+",
"1k",
"seed",
"dictionary",
"+",
"+",
"+",
"36.9",
"29.1",
"Table",
"4",
":",
"Model",
"Variation"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"B-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"For",
"consistency",
",",
"the",
"row",
"number",
"is",
"consistent",
"with",
"the",
"full",
"results",
"(",
"including",
"evaluation",
"onK",
",",
"warm",
"-",
"up",
",",
"tokenization",
",",
"and",
"t",
"%",
")",
"in",
"Appendix",
"B",
"All",
"the",
"models",
"are",
"based",
"on",
"the",
"XLM",
"instance",
"Row",
"2shows",
"the",
"default",
"configurations",
"we",
"use",
"in",
"UNMT"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName"
] |
[
"+",
"denotes",
"the",
"default",
"configuration",
"−denotes",
"an",
"inapplicable",
"term",
"UBWE",
"denotes",
"that",
"we",
"pre",
"-",
"train",
"the",
"bilingual",
"embeddings",
"unsupervisedly",
"and",
"then",
"pre",
"-",
"train",
"the",
"entire",
"model",
"with",
"our",
"method",
"In",
"1k",
"seed",
"dictionary",
"test",
",",
"the",
"model",
"employs",
"a",
"candidate",
"from",
"a",
"seed",
"dictionary",
"MUSE",
"score",
"XLM",
"(",
"Lample",
"and",
"Conneau",
",",
"2019",
")",
"0.55",
"+",
"[",
"C",
"]",
"x",
"0.61",
"MASS",
"(",
"Song",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricName",
"I-MetricName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
",",
"2019",
")",
"⋆",
"0.60",
"+",
"[",
"C",
"]",
"x",
"0.64",
"mBART",
"(",
"Liu",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"B-MetricValue",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"O",
"O",
"O",
"O",
"O"
] |
[
",",
"2020",
")",
"⋆",
"0.59",
"+",
"[",
"C",
"]",
"x",
"0.64",
"Table",
"5",
":",
"Performance",
"on",
"MUSE",
"task"
] | [
"O",
"O",
"O",
"O",
"B-MetricValue",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Baseline",
"models",
"(",
"⋆",
")",
"are",
"reimplemented",
"with",
"our",
"configurations",
"3.4",
"Cross",
"-",
"lingual",
"Word",
"Similarity",
"Setup",
"Given",
"the",
"idea",
"of",
"our",
"method",
",",
"we",
"consider",
"cross",
"-",
"lingual",
"mappings",
"of",
"tokens",
"Therefore",
",",
"we",
"are",
"interested",
"in",
"the",
"isomorphism",
"of",
"languages",
"’",
"embedding",
"spaces",
"To",
"further",
"investigate",
",",
"the",
"pre-",
"trained",
"UNMT",
"model",
"is",
"evaluated",
"on",
"MUSE",
"⋄",
"(",
"Lample",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-DatasetName",
"O",
"O",
"O",
"O",
"O"
] |
[
",",
"2018b",
")",
"with",
"the",
"provided",
"test",
"sets",
"and",
"tools",
",",
"which",
"is",
"used",
"to",
"test",
"cross",
"-",
"lingual",
"word",
"similarities",
"on",
"En↔De"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricName",
"I-MetricName",
"I-MetricName",
"I-MetricName",
"I-MetricName",
"O",
"O"
] |
[
"This",
"test",
"can",
"gener-",
"ally",
"evaluate",
"the",
"degree",
"of",
"the",
"isomorphism",
"of",
"languages",
"’",
"embedding",
"spaces",
"We",
"reuse",
"the",
"pre-",
"trained",
"models",
"in",
"our",
"UNMT",
"experiment",
"After",
"restoration",
",",
"we",
"extract",
"words",
"required",
"by",
"the",
"test",
"set",
"via",
"shared",
"lookup",
"tables",
"For",
"words",
"split",
"into",
"2",
"+",
"sub",
"-",
"tokens",
",",
"we",
"average",
"all",
"the",
"sub",
"-",
"tokens",
"Result",
"We",
"evaluate",
"the",
"performance",
"by",
"similar-",
"ities",
",",
"reporting",
"the",
"result",
"in",
"Table",
"5"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricName",
"I-MetricName",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Applying",
"[",
"C",
"]",
"xcan",
"increase",
"the",
"similarities",
"of",
"parallel",
"words",
"from",
"{",
"En",
",",
"De",
"}",
",",
"consistently",
"improving",
"the",
"per-",
"formance",
"of",
"the",
"models",
"on",
"this",
"task",
"It",
"indicates",
"that",
"[",
"C",
"]",
"xhelps",
"the",
"models",
"learn",
"token",
"-",
"level",
"cross-",
"linguality",
"in",
"pre",
"-",
"training",
"3.5",
"Cross",
"-",
"lingual",
"Classification",
"Setup",
"Beyond",
"UNMT",
"tasks",
"or",
"translation",
"tasks",
",",
"CLPM",
"can",
"consistently",
"improve",
"cross",
"-",
"lingual",
"transfer"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"B-TaskName",
"I-TaskName",
"O",
"B-TaskName",
"I-TaskName",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Then",
",",
"we",
"attempt",
"the",
"cross",
"-",
"lingual",
"clas-",
"sification",
"task",
"on",
"XNLI",
"(",
"Conneau",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"B-DatasetName",
"O",
"O",
"O",
"O"
] |
[
",",
"2018",
")",
"to",
"test",
"general",
"cross",
"-",
"linguality",
"[",
"C",
"]",
"ximproves",
"For",
"thisModel",
"Avg",
"(",
"Acc"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricName",
"I-MetricName",
"I-MetricName"
] |
[
")",
"mBERT",
"baseline",
"(",
"Wu",
"and",
"Dredze",
",",
"2019",
")",
"66.3",
"XLM",
"(",
"Lample",
"and",
"Conneau",
",",
"2019",
")",
"71.5",
"+",
"word",
"translation",
"tables",
"(",
"Chaudhary",
"et",
"al"
] | [
"I-MetricName",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
",",
"2020",
")",
"72.7",
"+",
"[",
"C",
"]",
"x",
"74.0",
"+",
"MT",
"(",
"Lample",
"and",
"Conneau",
",",
"2019",
")",
"75.1",
"Table",
"6",
":",
"Performance",
"of",
"cross",
"-",
"lingual",
"classification",
"on",
"XNLI"
] | [
"O",
"O",
"O",
"B-MetricValue",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MetricValue",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"B-DatasetName"
] |
[
"MT",
"stands",
"for",
"additional",
"parallel",
"corpora",
"We",
"use",
"the",
"same",
"transformer",
"models",
",",
"BPE",
"size",
",",
"corpora",
",",
"tokenization",
",",
"and",
"BLEU",
"as",
"the",
"baseline",
"models",
"(",
"see",
"more",
"details",
"in",
"Appendix",
"D.3",
")"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"test",
",",
"we",
"follow",
"the",
"standard",
"and",
"basic",
"experiment",
"(",
"Lample",
"and",
"Conneau",
",",
"2019",
")",
"to",
"train",
"a",
"12",
"-",
"layer",
"Transformer",
"encoder",
"with",
"80k",
"BPE",
"on",
"Wikipedia",
"dumps⋄of",
"15",
"XNLI",
"languages"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-HyperparameterValue",
"B-HyperparameterName",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"To",
"pre",
"-",
"train",
"the",
"encoder",
"on",
"Encorpora",
",",
"considering",
"the",
"zero",
"-",
"shot",
"classification",
"based",
"on",
"finetuning",
"EnNLI",
"dataset",
",",
"we",
"randomly",
"compute",
"[",
"C",
"]",
"xfrom",
"other",
"languages",
"with",
"equal",
"probability",
"to",
"avoid",
"the",
"cross",
"-",
"lingual",
"bias"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"O",
"B-DatasetName",
"I-DatasetName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"For",
"pre",
"-",
"training",
"on",
"corpora",
"of",
"other",
"lan-",
"guages",
",",
"we",
"only",
"compute",
"[",
"C",
"]",
"xin",
"the",
"Enentries",
"Note",
"that",
",",
"although",
"we",
"have",
"different",
"strategies",
"of",
"[",
"C",
"]",
"xfor",
"the",
"languages",
",",
"we",
"still",
"concatenated",
"all",
"the",
"corpora",
"of",
"the",
"languages",
"for",
"joint",
"pre",
"-",
"training",
"Af-",
"ter",
"pre",
"-",
"training",
",",
"we",
"deploy",
"a",
"randomly",
"initialized",
"linear",
"classifier",
"and",
"finetune",
"the",
"encoder",
"and",
"the",
"linear",
"classifier",
"on",
"the",
"EnNLI",
"dataset",
"with",
"mini-",
"batch",
"size",
"16"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-DatasetName",
"O",
"O",
"B-HyperparameterName",
"I-HyperparameterName",
"I-HyperparameterName",
"B-HyperparameterValue"
] |
[
"We",
"make",
"zero",
"-",
"shot",
"classifications",
"for",
"other",
"languages"
] | [
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"O"
] |
[
"See",
"more",
"details",
"in",
"Appendix",
"D.3",
"Result",
"We",
"report",
"the",
"result",
"in",
"Table",
"6",
"CLPM",
"shows",
"effectiveness",
"on",
"this",
"task",
",",
"outperforming",
"baseline",
"models"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"It",
"indicates",
"that",
"[",
"C",
"]",
"xcan",
"improve",
"cross",
"-",
"lingual",
"transfer",
"Meanwhile",
",",
"[",
"C",
"]",
"xunderper-",
"forms",
"XLM",
"+",
"MT",
"that",
"uses",
"parallel",
"corpora",
"to",
"improve",
"cross",
"-",
"linguality",
"As",
"discussed",
"earlier",
",",
"[",
"C",
"]",
"x",
"can",
"provide",
"token",
"-",
"level",
"cross",
"-",
"lingual",
"knowledge",
"at",
"the",
"very",
"least",
"but",
"is",
"less",
"effective",
"than",
"golden",
"sentence",
"-",
"level",
"knowledge",
"Although",
"XLM",
"+",
"MT862",
"uses",
"additional",
"datasets",
",",
"it",
"somewhat",
"sets",
"an",
"upper",
"bound"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"I-MethodName",
"I-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"On",
"the",
"other",
"hand",
",",
"our",
"method",
"outperforms",
"dictionary",
"-",
"based",
"methods",
"(",
"+",
"word",
"translation",
"ta-",
"bles",
")",
"Similar",
"to",
"the",
"observation",
"in",
"UNMT",
",",
"we",
"attribute",
"to",
"the",
"effectiveness",
"of",
"using",
"multiple",
"can-",
"didates",
"to",
"capture",
"morphological",
"variations"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"How-",
"ever",
",",
"to",
"avoid",
"cross",
"-",
"lingual",
"bias",
",",
"we",
"use",
"Enas",
"a",
"pivot",
"or",
"anchor",
"point",
"This",
"could",
"be",
"a",
"potential",
"problem",
"for",
"further",
"adaptation",
"to",
"other",
"multilingual",
"tasks",
"See",
"limitations",
"in",
"Appendix",
"A",
"4",
"Related",
"Work",
"and",
"Comparison",
"(",
"Ren",
"et",
"al",
",",
"2019a",
";",
"Chaudhary",
"et",
"al",
",",
"2020",
";",
"Lam-",
"ple",
"et",
"al",
",",
"2018c",
")",
"leverage",
"translation",
"tables",
"as",
"entries",
"for",
"the",
"other",
"languages",
",",
"which",
"are",
"auto-",
"matically",
"generated",
"from",
"statistical",
"models",
",",
"e.g.",
",",
"n",
"-",
"gram",
"models",
"The",
"model",
"forms",
"an",
"explicit",
"cross-",
"lingual",
"forward",
"pass",
":",
"{",
"[",
"M",
"]",
",",
"xj\\i",
"}",
"→ti",
",",
"where",
"ti",
"is",
"the",
"entry",
"of",
"the",
"other",
"language",
"for",
"xi",
"In",
"con-",
"trast",
",",
"our",
"method",
"has",
"two",
"significant",
"differences",
":",
"1",
")",
"we",
"focus",
"on",
"the",
"left",
"side",
",",
"adapting",
"our",
"[",
"C",
"]",
"xto",
"the",
"inputs",
"of",
"MLM",
";",
"2",
")",
"our",
"method",
"does",
"not",
"rely",
"on",
"token",
"/",
"phrase",
"-",
"level",
"translation",
"tables",
"Dufter",
"and",
"Schütze",
"(",
"2020",
")",
"present",
"a",
"cross",
"-",
"lingual",
"forward",
"pass",
":",
"{",
"nn",
",",
"x",
"j\\i",
"}",
"→",
"xi",
",",
"where",
"nnisxi",
"’s",
"near-",
"est",
"neighbor",
"of",
"the",
"other",
"language",
"in",
"the",
"space",
"of",
"UBWE",
"However",
",",
"UBWE",
"is",
"static",
"and",
"fixed",
"with-",
"out",
"any",
"interaction",
"with",
"the",
"multilingual",
"model"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"It",
"might",
"limit",
"what",
"it",
"can",
"be",
"ultimately",
"used",
"for",
"translation",
"(",
"Sun",
"et",
"al",
",",
"2019",
";",
"Artetxe",
"et",
"al",
",",
"2018b",
";",
"Lample",
"et",
"al",
",",
"2018a",
")",
"We",
"present",
"a",
"dynamic",
"ap-",
"proach",
"to",
"obtain",
"candidates",
"of",
"the",
"other",
"language",
"from",
"the",
"model",
"itself",
",",
"which",
"is",
"inspired",
"by",
"(",
"Ai",
"and",
"Fang",
",",
"2021b",
";",
"Sennrich",
"et",
"al",
",",
"2016a",
")",
"The",
"benefit",
"is",
"that",
"embeddings",
"and",
"representations",
"are",
"contex-",
"tualized",
"when",
"pre",
"-",
"training",
"MLM",
"on",
"monolingual",
"corpora",
"in",
"different",
"languages",
"(",
"Lample",
"and",
"Con-",
"neau",
",",
"2019",
")",
"Although",
"it",
"is",
"not",
"reliable",
"at",
"the",
"very",
"early",
"pre",
"-",
"training",
",",
"we",
"provide",
"a",
"compromised",
"ini-",
"tialization",
"for",
"this",
"problem",
"We",
"also",
"consider",
"multi-",
"ple",
"candidates",
"for",
"cross",
"-",
"lingual",
"prototypes",
"instead",
"ofnn",
",",
"which",
"is",
"softer",
"and",
"can",
"cover",
"morpholog-",
"ical",
"or",
"relevant",
"variations",
"in",
"the",
"other",
"language",
"On",
"the",
"other",
"hand",
",",
"considering",
"cross",
"-",
"lingual",
"pro-",
"totypes",
"is",
"not",
"a",
"novel",
"idea",
"for",
"cross",
"-",
"linguality",
",",
"(",
"Wang",
"et",
"al",
",",
"2019",
";",
"Huang",
"et",
"al",
",",
"2019",
";",
"Ai",
"and",
"Fang",
",",
"2021a",
")",
"present",
"methods",
"to",
"leverage",
"cross-",
"lingual",
"prototypes",
"to",
"guide",
"encoding",
"and",
"decoding",
",",
"forming",
"a",
"cross",
"-",
"lingual",
"forward",
"pass",
"by",
"modifying",
"inner",
"representations",
"of",
"encoding",
"and",
"decoding",
":",
"{",
"[",
"M",
"]",
",",
"xj\\i",
"}",
"→",
"{",
"[",
"M",
"]",
",",
"hxj",
",",
"hyi",
"}",
"→xi",
",",
"where",
"hyiis",
"an",
"approximation",
"of",
"xi",
"’s",
"inner",
"representation",
"in",
"encoding",
"and",
"decoding",
"from",
"the",
"other",
"language",
"It",
"results",
"in",
"a",
"different",
"direction",
"We",
"also",
"employ",
"the",
"alternation",
"strategy",
"that",
"can",
"be",
"viewed",
"as",
"linguistic",
"code",
"-",
"switching",
"(",
"Scotton",
"and",
"Ury",
",",
"1977",
")",
"somewhat",
",",
"where",
"the",
"model",
"is",
"pre",
"-",
"trained",
"in",
"more",
"linguistic",
"varieties",
"In",
"learn-",
"ing",
"models",
",",
"linguistic",
"code",
"-",
"switching",
"performs",
"as",
"data",
"augmentation",
"processes",
"(",
"Krishnan",
"et",
"al",
",",
"2021",
";",
"Chaudhary",
"et",
"al",
",",
"2020",
";",
"Tarunesh",
"et",
"al",
",",
"2021",
")",
"with",
"the",
"help",
"of",
"static",
"translation",
"tables",
"or",
"lexicon",
"induction",
"in",
"supervised",
"manners"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"O",
"O",
"O"
] |
[
"How-",
"ever",
",",
"lexicon",
"induction",
"datasets",
"or",
"translation",
"ta-",
"bles",
"have",
"been",
"reported",
"to",
"misrepresent",
"morpho-",
"logical",
"variations",
"and",
"overly",
"focus",
"on",
"named",
"en-",
"tities",
"and",
"frequent",
"words",
"(",
"Artetxe",
"et",
"al"
] | [
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
",",
"2020",
";",
"Czarnowska",
"et",
"al",
",",
"2020",
";",
"Kementchedjhieva",
"et",
"al",
",",
"2020",
")",
"In",
"contrast",
",",
"CLPM",
"is",
"dynamic",
"and",
"un-",
"supervised",
",",
"leveraging",
"contextualized",
"representa-",
"tions",
"and",
"multiple",
"morphological",
"variations",
"in",
"the",
"model",
"’s",
"embedding",
"space"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Meanwhile",
",",
"translation",
"tables",
"are",
"instances",
"/",
"embeddings",
"in",
"the",
"embedding",
"space",
",",
"whereas",
"cross",
"-",
"lingual",
"prototypes",
"do",
"not",
"ex-",
"ist",
"in",
"the",
"embedding",
"spaces",
"and",
"are",
"new",
"gener-",
"alized",
"samples",
"for",
"the",
"model",
"This",
"distinction",
"is",
"observed",
"from",
"the",
"discriminator",
"in",
"Figure",
"1",
"5",
"Conclusion",
"In",
"this",
"work",
",",
"we",
"present",
"CLPM",
",",
"an",
"alternative",
"masking",
"scheme",
",",
"to",
"compute",
"special",
"tokens",
"[",
"C",
"]",
"xfor",
"masking",
"in",
"multilingual",
"MLM",
"pre",
"-",
"training"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName"
] |
[
"[",
"C",
"]",
"xis",
"the",
"cross",
"-",
"lingual",
"prototype",
"for",
"the",
"selected",
"word",
"x",
",",
"computed",
"from",
"multiple",
"candidates",
"dynamically",
"and",
"token",
"-",
"wise",
"Compared",
"to",
"the",
"standard",
"masking",
"scheme",
"[",
"M",
"]",
",",
"[",
"C",
"]",
"xautomatically",
"forms",
"an",
"explicit",
"cross",
"-",
"lingual",
"forward",
"pass",
"in",
"attention",
"mechanism",
",",
"consistently",
"improving",
"cross",
"-",
"linguality",
"in",
"multilin-",
"gual",
"MLM",
"pre",
"-",
"training"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName"
] |
[
"Experiments",
"show",
"that",
"CLPM",
"can",
"consistently",
"improve",
"the",
"performance",
"of",
"translation",
"and",
"cross",
"-",
"lingual",
"transfer"
] | [
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName"
] |
[
"References",
"Martin",
"Abadi",
",",
"Paul",
"Barham",
",",
"Jianmin",
"Chen",
",",
"Zhifeng",
"Chen",
",",
"Andy",
"Davis",
",",
"Jeffrey",
"Dean",
",",
"Matthieu",
"Devin",
",",
"Sanjay",
"Ghemawat",
",",
"Geoffrey",
"Irving",
",",
"Michael",
"Isard",
",",
"Manjunath",
"Kudlur",
",",
"Josh",
"Levenberg",
",",
"Rajat",
"Monga",
",",
"Sherry",
"Moore",
",",
"Derek",
"G",
"Murray",
",",
"Benoit",
"Steiner",
",",
"Paul",
"Tucker",
",",
"Vijay",
"Vasudevan",
",",
"Pete",
"Warden",
",",
"Martin",
"Wicke",
",",
"Yuan",
"Yu",
",",
"and",
"Xiaoqiang",
"Zheng",
"2016",
"Ten-",
"sorflow",
":",
"A",
"system",
"for",
"large",
"-",
"scale",
"machine",
"learning"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Proceedings",
"of",
"the",
"61st",
"Annual",
"Meeting",
"of",
"the",
"Association",
"for",
"Computational",
"Linguistics",
"Volume",
"1",
":",
"Long",
"Papers",
",",
"pages",
"877–892",
"July",
"9",
"-",
"14",
",",
"2023",
"©",
"2023",
"Association",
"for",
"Computational",
"Linguistics",
"How",
"About",
"Kind",
"of",
"Generating",
"Hedges",
"using",
"End",
"-",
"to",
"-",
"End",
"Neural",
"Models",
"?",
"Alafate",
"Abulimiti1,2",
",",
"Chloé",
"Clavel3",
",",
"Justine",
"Cassell1,4",
"1INRIA",
",",
"Paris2ENS",
"/",
"PSL",
"<",
"alafate.abulimiti",
"@",
"inria.fr",
">",
"3LTCI",
",",
"Insitut",
"Polytechnique",
"de",
"Paris",
",",
"Telecom",
"Paris",
"<",
"chloe.clavel",
"@",
"telecom-paris.fr",
">",
"4Carnegie",
"Mellon",
"University",
"<",
"justine",
"@",
"cs.cmu.edu",
">",
"Abstract",
"Hedging",
"is",
"a",
"strategy",
"for",
"softening",
"the",
"impact",
"of",
"a",
"statement",
"in",
"conversation"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"In",
"reducing",
"the",
"strength",
"of",
"an",
"expression",
",",
"it",
"may",
"help",
"to",
"avoid",
"embarrassment",
"(",
"more",
"technically",
",",
"“",
"face",
"threat",
"”",
")",
"to",
"one",
"’s",
"listener",
"For",
"this",
"reason",
",",
"it",
"is",
"often",
"found",
"in",
"contexts",
"of",
"instruction",
",",
"such",
"as",
"tutoring",
"In",
"this",
"work",
",",
"we",
"develop",
"a",
"model",
"of",
"hedge",
"generation",
"based",
"on",
"i",
")",
"fine",
"-",
"tuning",
"state-",
"of",
"-",
"the",
"-",
"art",
"language",
"models",
"trained",
"on",
"human-",
"human",
"tutoring",
"data",
",",
"followed",
"by",
"ii",
")",
"reranking",
"to",
"select",
"the",
"candidate",
"that",
"best",
"matches",
"the",
"ex-",
"pected",
"hedging",
"strategy",
"within",
"a",
"candidate",
"pool",
"using",
"a",
"hedge",
"classifier"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"We",
"apply",
"this",
"method",
"to",
"a",
"natural",
"peer",
"-",
"tutoring",
"corpus",
"containing",
"a",
"significant",
"number",
"of",
"disfluencies",
",",
"repetitions",
",",
"and",
"repairs"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"B-DatasetName",
"I-DatasetName",
"I-DatasetName",
"I-DatasetName",
"I-DatasetName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"The",
"results",
"show",
"that",
"generation",
"in",
"this",
"noisy",
"environment",
"is",
"feasible",
"with",
"rerank-",
"ing",
"By",
"conducting",
"an",
"error",
"analysis",
"for",
"both",
"approaches",
",",
"we",
"reveal",
"the",
"challenges",
"faced",
"by",
"systems",
"attempting",
"to",
"accomplish",
"both",
"social",
"and",
"task",
"-",
"oriented",
"goals",
"in",
"conversation",
"1",
"Introduction",
"When",
"people",
"interact",
",",
"they",
"attend",
"not",
"just",
"to",
"the",
"task",
"at",
"hand",
",",
"but",
"also",
"to",
"their",
"relationship",
"with",
"their",
"interlocutors",
"(",
"Tracy",
"and",
"Coupland",
",",
"1990",
")",
"One",
"key",
"aspect",
"of",
"the",
"relationship",
"that",
"people",
"attend",
"to",
",",
"while",
"engaging",
"in",
"contexts",
"as",
"diverse",
"as",
"sales",
"(",
"Gremler",
"and",
"Gwinner",
",",
"2008",
";",
"Planken",
",",
"2005",
")",
",",
"education",
"(",
"Glazier",
",",
"2016",
";",
"Murphy",
"and",
"Rodríguez",
"-",
"Manzanares",
",",
"2012",
")",
"and",
"healthcare",
"(",
"Di-",
"Matteo",
",",
"1979",
";",
"Leach",
",",
"2005",
")",
",",
"is",
"what",
"is",
"referred",
"to",
"as",
"rapport",
",",
"a",
"sense",
"of",
"harmony",
"and",
"mutual",
"un-",
"derstanding",
"between",
"participants",
"in",
"a",
"conversation",
"(",
"Spencer",
"-",
"Oatey",
",",
"2005",
";",
"Tickle",
"-",
"Degnen",
"and",
"Rosen-",
"thal",
",",
"1990",
")",
"Indeed",
",",
"higher",
"levels",
"of",
"rapport",
"are",
"cor-",
"related",
"with",
"better",
"performance",
"in",
"each",
"of",
"these",
"do-",
"mains",
"Zhao",
"et",
"al",
"(",
"2014",
")",
"describes",
"rapport",
"as",
"built",
"upon",
"a",
"base",
"of",
"mutual",
"attentiveness",
",",
"face",
"manage-",
"ment",
",",
"and",
"coordination",
"This",
"base",
"is",
"built",
"primarily",
"by",
"conversational",
"strategies",
",",
"or",
"ways",
"of",
"speaking",
"(",
"including",
"nonverbal",
"and",
"paraverbal",
"behaviors",
")",
"that",
"yeah",
",",
"so",
"it",
"will",
"be",
"...",
"I",
"think",
"you",
"could",
"add",
"four",
"to",
"both",
"sides",
"oh",
"no",
"...",
"Figure",
"1",
":",
"Hedging",
"in",
"peer",
"tutoring",
"manage",
"rapport",
"throughout",
"a",
"conversation"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Key",
"conversational",
"strategies",
"include",
"self",
"-",
"disclosure",
",",
"reference",
"to",
"shared",
"experience",
",",
"praise",
",",
"and",
"hedging",
"—",
"giving",
"instructions",
"or",
"conveying",
"information",
"in",
"an",
"indirect",
"manner",
"when",
"it",
"might",
"otherwise",
"sound",
"rude",
"or",
"overly",
"demanding"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"End",
"-",
"to",
"-",
"end",
"large",
"language",
"models",
"(",
"LLM",
")",
",",
"of",
"the",
"kind",
"that",
"are",
"increasingly",
"popular",
"and",
"powerful",
",",
"do",
"a",
"good",
"job",
"at",
"carrying",
"out",
"the",
"propositional",
"or",
"information",
"-",
"carrying",
"aspects",
"of",
"conversation",
",",
"and",
"a",
"relatively",
"good",
"job",
"of",
"maintaining",
"the",
"coherence",
"of",
"a",
"conversation",
",",
"but",
"they",
"are",
"not",
"as",
"good",
"at",
"changing",
"how",
"they",
"say",
"something",
"as",
"a",
"function",
"of",
"a",
"relation-",
"ship",
"with",
"the",
"human",
"user",
",",
"while",
"humans",
"are",
",",
"for",
"the",
"most",
"part",
",",
"quite",
"good",
"at",
"this"
] | [
"B-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"However",
",",
"since",
"saying",
"things",
"in",
"a",
"specific",
"manner",
"-",
"for",
"example",
",",
"through",
"a",
"hedge",
"-",
"helps",
"task",
"performance",
",",
"it",
"is",
"an",
"important",
"topic",
"for",
"dialogue",
"systems"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Linguists",
"define",
"hedges",
"as",
"a",
"way",
"of",
"diminishing",
"face",
"threat",
"(",
"meaning",
"the",
"“",
"positive",
"social",
"value",
"a",
"person",
"effectively",
"claims",
"for",
"himself",
"”",
"(",
"Goffman",
",",
"1967",
")",
"by",
"attenuating",
"the",
"extent",
"or",
"impact",
"of",
"an",
"ex-",
"pression",
"(",
"Brown",
"and",
"Levinson",
",",
"1987",
";",
"Fraser",
",",
"2010",
")",
"Figure",
"1",
"shows",
"a",
"typical",
"example",
"of",
"hedging",
"in",
"a",
"peer",
"tutoring",
"setting",
",",
"where",
"the",
"tutor",
"uses",
"two",
"hedges",
"(",
"“",
"I",
"think",
"”",
"and",
"“",
"could",
"”",
"rather",
"than",
"“",
"should",
"”",
")",
"to",
"deliver",
"a",
"hint",
"for",
"the",
"next",
"step",
"of",
"solving",
"an",
"algebra",
"equation"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Tutoring",
"is",
"one",
"context",
"in",
"which",
"hedges",
"are",
"found",
"in",
"abundance",
"and",
"where",
"recognizing",
"them",
"might",
"be",
"important",
"for",
"intelligent",
"tutoring",
"systems",
",",
"as",
"attested",
"by",
"the",
"number",
"of",
"computational",
"ap-877",
"proaches",
"that",
"attempt",
"to",
"do",
"so",
"(",
"see",
"section",
"2",
")"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"In-",
"terestingly",
",",
"even",
"unskilled",
"tutors",
"use",
"them",
"In",
"fact",
",",
"research",
"on",
"peer",
"tutoring",
"has",
"shown",
"that",
"when",
"rap-",
"port",
"between",
"a",
"peer",
"tutor",
"and",
"tutee",
"is",
"low",
",",
"but",
"the",
"tutor",
"is",
"confident",
"in",
"his",
"/",
"her",
"skills",
",",
"that",
"tutor",
"tends",
"to",
"use",
"more",
"hedges",
",",
"and",
"this",
"results",
"in",
"more",
"prob-",
"lems",
"attempted",
"by",
"the",
"student",
"and",
"more",
"problems",
"successfully",
"solved",
"(",
"Madaio",
"et",
"al",
",",
"2017",
")",
"In",
"this",
"paper",
",",
"then",
",",
"we",
"work",
"towards",
"the",
"devel-",
"opment",
"of",
"a",
"generation",
"module",
"for",
"a",
"virtual",
"peer",
"tutor",
"that",
",",
"like",
"real",
"peer",
"tutors",
",",
"is",
"able",
"to",
"choose",
"the",
"manner",
"of",
"delivering",
"information",
"in",
"such",
"a",
"way",
"Specifically",
",",
"we",
"address",
"two",
"research",
"questions",
":",
"RQ1",
":",
"How",
"good",
"are",
"end",
"-",
"to",
"-",
"end",
"large",
"language",
"models",
"used",
"alone",
"for",
"generating",
"hedges",
"when",
"fine-",
"tuned",
"on",
"a",
"peer",
"-",
"tutoring",
"dialogue",
"dataset",
"?"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Are",
"the",
"models",
"able",
"to",
"implicitly",
"learn",
"when",
"and",
"how",
"to",
"generate",
"hedges",
"?"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"O"
] |
[
"The",
"first",
"question",
"may",
"be",
"answered",
"by",
"compar-",
"ing",
"the",
"performance",
"of",
"various",
"fine",
"-",
"tuned",
"models",
"If",
"the",
"end",
"-",
"to",
"-",
"end",
"models",
"can",
"not",
"learn",
"to",
"hedge",
"im-",
"plicitly",
",",
"we",
"might",
"attempt",
"to",
"drive",
"the",
"models",
"to",
"generate",
"the",
"utterances",
"by",
"providing",
"the",
"correct",
"labels"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"We",
"assume",
"that",
"the",
"correct",
"labels",
"can",
"be",
"provided",
"by",
"another",
"module",
"of",
"the",
"system",
",",
"so",
"we",
"compare",
"the",
"reranking",
"method",
"with",
"the",
"fine",
"-",
"tuning",
"method",
",",
"as",
"the",
"former",
"is",
"simple",
",",
"powerful",
",",
"and",
"widely",
"used",
"for",
"text",
"generation"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"I-MethodName",
"O",
"O",
"B-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Consequently",
",",
"the",
"second",
"question",
"is",
":",
"RQ2",
":",
"Can",
"we",
"improve",
"these",
"models",
"by",
"using",
"a",
"reranking",
"approach",
"?"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"I-MethodName",
"O"
] |
[
"If",
"so",
",",
"what",
"are",
"the",
"remaining",
"errors",
"and",
"why",
"do",
"they",
"occur",
"?",
"2",
"Related",
"Work",
"Considerably",
"more",
"computational",
"methods",
"exist",
"to",
"determine",
"what",
"a",
"dialogue",
"system",
"should",
"say",
"than",
"how",
"to",
"say",
"it",
"However",
",",
"more",
"recently",
",",
"with",
"the",
"increased",
"power",
"of",
"end",
"-",
"to",
"-",
"end",
"models",
"to",
"find",
"information",
"and",
"convey",
"it",
"accurately",
",",
"we",
"can",
"now",
"turn",
"to",
"ensuring",
"that",
"the",
"end",
"-",
"to",
"-",
"end",
"model",
"simul-",
"taneously",
"also",
"meets",
"social",
"goals",
",",
"to",
"increase",
"the",
"impact",
"and",
"acceptability",
"of",
"what",
"is",
"conveyed",
"2.1",
"Theoretical",
"Approaches",
"to",
"hedges",
"As",
"described",
"above",
",",
"a",
"hedge",
"can",
"soften",
"the",
"impact",
"of",
"an",
"utterance",
"that",
"might",
"otherwise",
"seem",
"rude",
",",
"such",
"as",
"a",
"demand",
"(",
"“",
"could",
"you",
"pass",
"the",
"salt",
"”",
")",
"or",
"an",
"instruction",
"(",
"“",
"you",
"might",
"want",
"to",
"pour",
"the",
"coffee",
"over",
"the",
"sink",
"”",
")"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Madaio",
"et",
"al",
"(",
"2017",
")",
"has",
"attested",
"to",
"the",
"frequent",
"use",
"of",
"hedges",
"in",
"the",
"peer",
"-",
"tutoring",
"setting",
",",
"and",
"their",
"positive",
"impact",
"on",
"performance",
",",
"perhaps",
"because",
"hedges",
"in",
"this",
"context",
"might",
"re-",
"duce",
"a",
"tutee",
"’s",
"embarrassment",
"at",
"not",
"knowing",
"the",
"correct",
"answer",
"(",
"Rowland",
",",
"2007",
")",
"In",
"linguistic",
"terms",
",",
"hedging",
"is",
"a",
"rhetorical",
"strat-",
"egy",
"that",
"attenuates",
"the",
"full",
"force",
"of",
"an",
"expression",
"(",
"Fraser",
",",
"2010",
")",
"and",
"for",
"this",
"reason",
",",
"it",
"has",
"been",
"cov-",
"ered",
"in",
"linguistic",
"pragmatics",
"and",
"the",
"study",
"of",
"polite-",
"ness",
"Two",
"main",
"categories",
"of",
"hedges",
"are",
"identified",
"in",
"the",
"literature",
":",
"Propositional",
"Hedges",
"andRela-",
"tional",
"Hedges",
"(",
"Prince",
"et",
"al",
",",
"1982",
")",
"Propositional",
"Hedges",
"(",
"called",
"Approximators",
"by",
"(",
"Prince",
"et",
"al",
",",
"1982",
")",
")",
"refer",
"to",
"uncertain",
"(",
"Vincze",
",",
"2014",
")",
",",
"fuzzy",
"(",
"Lakoff",
",",
"1975",
")",
"and",
"vague",
"(",
"Williamson",
",",
"2002",
")",
"lan-",
"guage",
"use",
",",
"such",
"as",
"“",
"kind",
"of",
"”",
"Relational",
"Hedges",
"(",
"called",
"Shields",
"in",
"(",
"Prince",
"et",
"al",
",",
"1982",
")",
")",
"indicate",
"that",
"the",
"expression",
"is",
"subjective",
"or",
"an",
"opinion",
",",
"as",
"in",
"“",
"I",
"think",
"that",
"is",
"incorrect",
"”",
"Attribution",
"Shields",
"are",
"a",
"subtype",
"of",
"relational",
"hedges",
"that",
"attribute",
"the",
"opinion",
"to",
"others",
",",
"such",
"as",
"“",
"everyone",
"says",
"you",
"should",
"stop",
"smoking",
"”",
"Apologizers",
"(",
"Raphalen",
"et",
"al",
",",
"2022",
")",
"are",
"apologies",
"that",
"mitigate",
"the",
"strength",
"of",
"an",
"utterance",
",",
"as",
"in",
"“",
"I",
"’m",
"sorry",
"but",
"you",
"have",
"to",
"do",
"your",
"homework",
"”",
"While",
"the",
"different",
"types",
"of",
"hedges",
"operate",
"in",
"different",
"ways",
",",
"they",
"all",
"serve",
"the",
"same",
"mitigation",
"functions",
"in",
"conversation",
"For",
"this",
"reason",
",",
"in",
"what",
"follows",
"—",
"a",
"first",
"attempt",
"at",
"generating",
"hedges",
"—",
"we",
"collapse",
"the",
"different",
"sub",
"-",
"classes",
"and",
"refer",
"only",
"to",
"hedges",
"and",
"non",
"-",
"hedges"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"2.2",
"Computational",
"Approaches",
"Some",
"prior",
"work",
"has",
"looked",
"at",
"the",
"detection",
"of",
"conversational",
"strategies",
"and",
"in",
"particular",
"work",
"by",
"Zhao",
"and",
"colleagues",
"(",
"Zhao",
"et",
"al",
",",
"2014",
",",
"2016b",
",",
"a",
")",
"Madaio",
"et",
"al",
"(",
"2017",
")",
"built",
"a",
"classifier",
"to",
"detect",
"hedg-",
"ing",
"and",
"achieved",
"an",
"accuracy",
"of",
"81",
"%"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"O",
"B-MetricName",
"O",
"B-MetricValue",
"I-MetricValue"
] |
[
"Recent",
"work",
"by",
"Raphalen",
"et",
"al",
"(",
"2022",
")",
"improved",
"the",
"detection",
"of",
"different",
"types",
"of",
"hedges",
"and",
"achieved",
"a",
"weighted",
"F1",
"score",
"of",
"0.97"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName",
"O",
"O",
"O",
"B-MetricName",
"I-MetricName",
"I-MetricName",
"O",
"B-MetricValue"
] |
[
"Hedging",
"is",
"a",
"particular",
"kind",
"of",
"indirectness",
",",
"and",
"therefore",
"as",
"we",
"look",
"at",
"prior",
"work",
"in",
"the",
"area",
",",
"we",
"include",
"approaches",
"to",
"the",
"generation",
"of",
"indi-",
"rect",
"speech"
] | [
"B-TaskName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"The",
"plan",
"-",
"based",
"generation",
"of",
"indirect",
"speech",
"acts",
"has",
"existed",
"almost",
"as",
"long",
"as",
"dialogue",
"systems",
"themselves",
"(",
"Clark",
",",
"1979",
";",
"Brown",
",",
"1980",
";",
"Perrault",
",",
"1980",
")",
"More",
"recently",
",",
"other",
"relevant",
"as-",
"pects",
"of",
"politeness",
"have",
"also",
"been",
"addressed",
"For",
"example",
",",
"Porayska",
"-",
"Pomsta",
"and",
"Mellish",
"(",
"2004",
")",
"op-",
"erationalized",
"the",
"important",
"notion",
"of",
"face",
"in",
"po-",
"liteness",
"theory",
"to",
"generate",
"polite",
"sentences",
"with",
"a",
"template",
"pool",
"Although",
"contemporary",
"dialogue878",
"systems",
"tend",
"to",
"integrate",
"indirect",
"speech",
"(",
"Miehle",
"et",
"al",
",",
"2022",
";",
"Briggs",
"et",
"al",
",",
"2017",
")",
",",
"generating",
"hedges",
"with",
"powerful",
"language",
"models",
",",
"and",
"particularly",
"as",
"a",
"function",
"of",
"the",
"social",
"context",
",",
"has",
"not",
"been",
"explored",
"Our",
"desire",
"to",
"look",
"at",
"the",
"social",
"context",
"leads",
"us",
"to",
"train",
"on",
"spontaneous",
"dialogue",
"that",
"is",
"substantially",
"noisier",
",",
"owing",
"to",
"natural",
"conversa-",
"tional",
"phenomena",
"such",
"as",
"disfluency",
"This",
"differs",
"from",
"the",
"majority",
"of",
"prior",
"work",
",",
"trained",
"on",
"written",
"or",
"acted",
"corpora",
"(",
"Li",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-DatasetName",
"I-DatasetName",
"I-DatasetName",
"I-DatasetName",
"O",
"O",
"O",
"O"
] |
[
",",
"2017",
";",
"Rashkin",
"et",
"al",
",",
"2019",
")",
"2.3",
"Generation",
"Techniques",
"Different",
"techniques",
"have",
"been",
"used",
"in",
"the",
"past",
"to",
"generate",
"responses",
"of",
"a",
"particular",
"kind",
"for",
"dialogue",
"systems",
"Madaan",
"et",
"al",
"(",
"2020",
")",
"used",
"n",
"-",
"gram",
"TF-",
"IDF",
"to",
"identify",
"source",
"style",
"words",
"and",
"generate",
"target",
"politeness",
"style",
"utterances",
"by",
"replacing",
"these",
"words"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"I-MethodName",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O"
] |
[
"Niu",
"and",
"Bansal",
"(",
"2018",
")",
"generated",
"politeness",
"formulations",
"by",
"using",
"reinforcement",
"learning",
"with",
"a",
"trained",
"politeness",
"classifier",
"Similar",
"to",
"our",
"ap-",
"proach",
",",
"the",
"explicit",
"knowledge",
"of",
"politeness",
"is",
"only",
"given",
"to",
"the",
"classifier",
"Liu",
"et",
"al",
"(",
"2021",
")",
"constructed",
"an",
"emotional",
"support",
"dataset",
"with",
"eight",
"different",
"dialogue",
"strategies",
"and",
"fine",
"-",
"tuned",
"the",
"pre",
"-",
"trained",
"language",
"models",
"by",
"connecting",
"the",
"label",
"tokens",
"to",
"the",
"beginning",
"of",
"each",
"utterance",
"in",
"order",
"to",
"create",
"a",
"dialogue",
"generator",
"that",
"can",
"produce",
"the",
"target",
"responses",
"without",
"focusing",
"on",
"the",
"social",
"context",
"The",
"reranking",
"method",
"is",
"also",
"widely",
"used",
"in",
"text",
"generation",
"tasks"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"I-MethodName",
"O",
"O",
"O",
"O",
"B-TaskName",
"I-TaskName",
"I-TaskName",
"I-TaskName"
] |
[
"Hossain",
"et",
"al",
"(",
"2020",
")",
"used",
"a",
"sim-",
"ple",
"and",
"effective",
"pipeline",
"where",
"they",
"retrieved",
"the",
"original",
"texts",
"from",
"the",
"database",
",",
"then",
"edited",
"with",
"a",
"Transformer",
"(",
"Vaswani",
"et",
"al"
] | [
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"O",
"B-MethodName",
"O",
"O",
"O",
"O"
] |
Subsets and Splits